Spatial Segmentation Radiotherapy Guidance Method and System Based on Target Immune Region Activation

Abstract

This application discloses a spatial segmentation radiotherapy guidance method and system based on target immune region activation, relating to the field of digital electrophysiological data processing. The method achieves three-dimensional visualization of metabolic activity and immune response through multimodal fusion of spectral CT / functional magnetic resonance imaging and ResNet radiomics feature extraction, which significantly improves the recognition accuracy compared to anatomical target volume. Furthermore, the immune thermogram generated by the XGBoost model and SAMed-2 segmentation markers can accurately locate highly immune-responsive voxels, and the Monte Carlo simulation gradient dose distribution ensures that the dose intensity in the immune-activated region is several times that of the conventional region while protecting the inactive region. The XGBoost model training of this application can dynamically adjust the dose hotspot according to the patient's specific immune characteristics, so that the radiation can accurately and sufficiently reach the region in the tumor area that is most capable of activating the immune system.

Description

Technical Field

[0001] This application relates to the field of electronic digital data processing technology, and in particular to a spatial segmentation radiotherapy guidance method and system based on target immune region activation. Background Technology

[0002] In current medical treatment, nearly 70% of malignant tumor patients receive radiotherapy. The current concept of radiotherapy is to ensure complete dose coverage of the entire tumor. Therefore, most radiotherapy equipment platforms currently utilize linear accelerators. Linear accelerators generate high-energy rays by accelerating electrons to "target" the tumor. The accelerator head is equipped with a multi-leaf grating, which consists of several pairs of 0.5–1 cm lead gates. Each lead gate is driven by a mechanism, resulting in a complex driving mechanism for the entire multi-leaf grating. Through the movement of these lead gates, the accelerator shapes the tumor, while the gantry rotates around the patient, ensuring that the output radiation beam covers the tumor 360°, achieving better dose conformal volumetric intensity-modulated radiotherapy (IMRT).

[0003] In the new study, the uneven distribution of radiation dose is beneficial for tumor control. This uneven dose distribution alters the immune microenvironment within the tumor, which is similar to activating the immune system, making the treatment more effective. Therefore, the non-uniform irradiation leads to improved efficacy.

[0004] Under current conditions, non-uniform irradiation primarily involves distributing a uniform dose within the tumor, a technique known as "lattice" radiotherapy, where each dose hotspot represents a lattice. However, because this method involves uniform distribution, the lattice is not the region within the tumor area most capable of activating the immune system, resulting in less than ideal therapeutic effects from current non-uniform irradiation treatments. Summary of the Invention

[0005] The main objective of this application is to provide a spatial segmentation radiotherapy guidance method and system based on target area immune region activation, in order to solve the problem that the non-uniform irradiation of "lattice" radiotherapy in the prior art is not the region in the tumor area that can most activate the immune system, resulting in the unsatisfactory treatment effect of the current non-uniform irradiation treatment method.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] A spatial segmentation radiotherapy guidance method based on target area immune region activation, the radiotherapy guidance method being applied to patients undergoing radiotherapy treatment using a pre-set radiotherapy instrument, the radiotherapy guidance method comprising:

[0008] Step S1: Obtain the spectral CT and functional magnetic resonance imaging data of the patient individual, and generate a fused image matrix dataset using a bilinear registration algorithm;

[0009] Step S2: Extract the radiomics features of the fused image matrix dataset using a ResNet convolutional network;

[0010] Step S3: Identify the metabolically active regions of the radiomics features using Otsu adaptive threshold segmentation to obtain an immune activity feature vector set;

[0011] Step S4: Train the XGBoost classification model using a preset dataset based on the XGBoost classification model, and input the immune activity feature vector set into the trained XGBoost classification model to obtain an immune heatmap;

[0012] Step S5: Obtain the radiotherapy localization CT of the patient and align it with the coordinate system of the immune thermogram. Then, use the SAMed-2 image segmentation model to perform voxel-level segmentation on the aligned immune thermogram and mark the voxels with high immune response.

[0013] Step S6: Obtain the three-dimensional coordinates of all immune high-response voxels and integrate them into a three-dimensional coordinate set of the immune target area. Iterate the motion trajectory of the multi-leaf grating of the preset radiotherapy instrument through Monte Carlo simulation, and cover all three-dimensional coordinates with the rays of the preset radiotherapy instrument in the form of a preset ray intensity gradient decreasing.

[0014] Step S7: Obtain the motion control command sequence after the multi-leaf grating motion trajectory iteration is completed and input it into the preset radiotherapy instrument.

[0015] As a further improvement to this application, step S7 involves obtaining the motion control command sequence after the multi-leaf grating motion trajectory iteration is completed and inputting it into the preset radiotherapy instrument. Following this, the process includes:

[0016] Step S10: After the radiotherapy treatment is completed by the preset radiotherapy instrument, obtain the SUVmax value matrix of the patient based on FDG-PET / CT scan;

[0017] Step S20: Calculate the target area FDG uptake change rate based on the baseline SUV and the SUVmax value matrix;

[0018] Step S30: Determine whether the change rate of FDG uptake in the target area exceeds a preset change rate threshold. If not, proceed to step S40.

[0019] Step S40: Determine that the motion control command sequence is invalid radiotherapy guidance;

[0020] In step S50, in response to the invalid radiotherapy guidance, steps S4 to S6 are repeated using an additional preset sample set as the execution subject to continue iteratively updating the XGBoost classification model;

[0021] Step S60: Obtain the motion control command sequence after further iteration and input it into the preset radiotherapy instrument;

[0022] Step S70: Repeat steps S10 to S60 until the change rate of FDG uptake in the target area exceeds the preset change rate threshold.

[0023] As a further improvement to this application, step S2, extracting the radiomics features of the fused image matrix dataset using a ResNet convolutional network, includes:

[0024] Step S21: Normalize the fused image matrix dataset;

[0025] Step S22: Divide the normalized fused image matrix dataset into cubes of a preset size;

[0026] Step S23: Define a ResNet convolutional network and define the number of channels in the input layer of the ResNet convolutional network to match the preset size;

[0027] Step S24: Input the fused image matrix dataset into the input layer of the ResNet convolutional network, and obtain a multidimensional deep feature vector set through forward propagation of the ResNet convolutional network;

[0028] Step S25: Filter the top feature vectors in the multidimensional deep feature vector set using the mRMR algorithm;

[0029] Step S26: All the top feature vectors are upsampled through a deconvolution layer to reconstruct a three-dimensional feature map, and the three-dimensional feature map is defined as the radiomics feature.

[0030] As a further improvement to this application, step S3 involves identifying metabolically active regions of the radiomics features using Otsu adaptive threshold segmentation to obtain an immune activity feature vector set, including:

[0031] Step S31: Normalize the grayscale of the three-dimensional feature map and calculate the histogram distribution of the three-dimensional feature map after grayscale normalization.

[0032] Step S32: Calculate the inter-class variance of the histogram distribution using the Otsu adaptive threshold;

[0033] Step S33: Select the optimal threshold among all candidate thresholds that maximizes the inter-class variance, and segment the three-dimensional feature map according to the optimal threshold to form high-metabolic regions and low-metabolic regions.

[0034] Step S34: Define the high-metabolic region as the metabolically active region;

[0035] Step S35: Perform morphological closing operation on the metabolically active region based on the preset structural voxel to obtain several connected regions.

[0036] Step S36: Calculate the gray-level co-occurrence matrix features for each connected region;

[0037] Step S37: Convert all gray-level co-occurrence matrix features into multi-dimensional feature vectors using a preset Python script;

[0038] Step S38: Integrate all multidimensional feature vectors and define them as the immune activity feature vector set.

[0039] As a further improvement to this application, step S4 involves training the XGBoost classification model using a preset dataset, and inputting the immune activity feature vector set into the trained XGBoost classification model to obtain an immune heatmap, including:

[0040] Step S41: Define the model parameters of the XGBoost classification model based on the preset model strategy;

[0041] Step S42: Standardize and reduce the dimensions of the preset dataset, and input the sample set in the preset dataset into the XGBoost classification model in batches for training;

[0042] Step S43: In each batch of training, the XGBoost classification model is validated using the validation set in the preset dataset, and training is terminated when the AUC value of the validation set no longer increases, thus obtaining the XGBoost classification model after training.

[0043] Step S44: Input the immune activity feature vector set into the trained XGBoost classification model and obtain the positive probability prediction value for each voxel.

[0044] Step S45: Calculate the numerical range of all positive probability prediction values, and assign the numerical range to at least three consecutive and different heat map color levels.

[0045] Step S46: Map all predicted positive probability values ​​and corresponding heatmap color levels to the coordinate system three-dimensional matrix of the fused image matrix dataset to obtain the immune heatmap.

[0046] As a further improvement to this application, step S5 involves acquiring the radiotherapy localization CT scan of the patient and aligning it with the immunothermal map coordinate system. The aligned immunothermal map is then segmented at the voxel level using the SAMed-2 image segmentation model, while simultaneously marking highly immune-responsive voxels, including:

[0047] Step S51: Obtain the centroid of the coordinate system of the radiotherapy positioning CT and the centroid of the coordinate system of the immunothermograph;

[0048] Step S52: Calculate the rotation and translation matrix between the radiotherapy positioning CT and the immunothermograph based on the centroids of the two coordinate systems;

[0049] Step S53: Based on the rotation and translation matrix, the radiotherapy positioning CT and the immunothermograph are registered using bicubic interpolation to achieve coordinate system alignment;

[0050] Step S54: The aligned immune heatmap is segmented at the voxel level using the SAMed-2 image segmentation model;

[0051] Step S55: Mark all voxels in the high-metabolic region as highly immune-responsive voxels.

[0052] As a further improvement to this application, step S6 involves obtaining the three-dimensional coordinates of all highly responsive immune voxels and integrating them into a three-dimensional coordinate set of the immune target area. The motion trajectory of the multi-leaf grating of the preset radiotherapy instrument is iterated using Monte Carlo simulation. The rays from the preset radiotherapy instrument are then used to cover all three-dimensional coordinates in a preset ray intensity gradient decreasing manner, including:

[0053] Step S61: Extract the three-dimensional coordinates of all highly immune voxels based on the immune heatmap segmented by the SAMed-2 image segmentation model.

[0054] Step S62: Generate the boundary surface based on all three-dimensional coordinates using the Delaunay triangulation algorithm;

[0055] Step S63: Using the center of the boundary surface as a reference point, set the dose attenuation coefficient according to a preset interval distance;

[0056] Step S64: Determine the initial opening and closing state of the multi-leaf grating and the initial angle of the collimator of the preset radiotherapy instrument based on the projection profile of the boundary surface.

[0057] Step S65: Based on the initial angle of the collimator, the optimal opening and closing combination of the multi-leaf grating at each angle of the collimator is calculated iteratively using a global optimization algorithm with a preset angle step size. All optimal opening and closing combinations are based on the projection contour that can just completely cover the boundary surface.

[0058] Step S66: Using the motion time sequence of the collimator as the chronological order, convert the position sequence of all optimal opening and closing combinations of the multileaf grating into the motion trajectory of the multileaf grating along the motion time sequence.

[0059] To achieve the above objectives, this application also provides the following technical solutions:

[0060] A spatial segmentation radiotherapy guidance system based on target area immune region activation, wherein the radiotherapy guidance system is applied to the radiotherapy guidance method described above, and the radiotherapy guidance system comprises:

[0061] The fused image matrix dataset acquisition module is used to acquire the spectral CT and functional magnetic resonance data of the patient individual and generate a fused image matrix dataset through a bilinear registration algorithm.

[0062] The radiomics feature extraction module is used to extract radiomics features from the fused image matrix dataset through a ResNet convolutional network.

[0063] The immune activity feature vector set acquisition module is used to identify the metabolically active regions of the radiomics features through Otsu adaptive threshold segmentation to obtain the immune activity feature vector set.

[0064] The immune heatmap acquisition module is used to train the XGBoost classification model based on the XGBoost classification model using a preset dataset, and input the immune activity feature vector set into the trained XGBoost classification model to obtain the immune heatmap.

[0065] The immune high-response voxel labeling module is used to acquire the radiotherapy positioning CT of the patient and align it with the coordinate system of the immune thermogram. The aligned immune thermogram is then segmented at the voxel level using the SAMed-2 image segmentation model, and immune high-response voxels are labeled simultaneously.

[0066] The multileaf grating motion trajectory iteration module is used to acquire the three-dimensional coordinates of all immune high-response voxels and integrate them into a three-dimensional coordinate set of the immune target area. Through Monte Carlo simulation, the multileaf grating motion trajectory of the preset radiotherapy instrument is iterated, and the rays of the preset radiotherapy instrument cover all three-dimensional coordinates in the form of a preset ray intensity gradient decreasing.

[0067] The motion control command sequence output module is used to acquire the motion control command sequence after the multileaf grating motion trajectory iteration is completed and input it into the preset radiotherapy instrument.

[0068] To achieve the above objectives, this application also provides the following technical solutions:

[0069] An electronic device includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the radiotherapy guidance method as described above.

[0070] To achieve the above objectives, this application also provides the following technical solutions:

[0071] A storage medium storing program instructions that, when executed by a processor, implement the radiotherapy guidance method described above.

[0072] Beneficial effects:

[0073] This application acquires spectral CT and functional magnetic resonance imaging (fMRI) data of individual patients and generates a fused image matrix dataset using a bilinear registration algorithm. It then extracts radiomics features from the fused image matrix dataset using a ResNet convolutional network. Otsu adaptive threshold segmentation identifies metabolically active regions of the radiomics features, yielding an immune activity feature vector set. An XGBoost classification model is trained using a pre-defined dataset, and the immune activity feature vector set is input into the trained XGBoost model to obtain an immune heatmap. The application acquires radiotherapy localization CT scans of individual patients and aligns them with the immune heatmap coordinate system. The aligned immune heatmap is then segmented at the voxel level using a SAMed-2 image segmentation model, while simultaneously marking highly responsive voxels. The three-dimensional coordinates of all highly responsive voxels are acquired and integrated into a three-dimensional coordinate set of the immune target area. Monte Carlo simulation iterates the motion trajectory of a pre-defined radiotherapy instrument's multi-leaf grating, ensuring that the radiation from the pre-defined radiotherapy instrument covers all three-dimensional coordinates in a decreasing gradient. Finally, the motion control command sequence after the multi-leaf grating motion trajectory iteration is acquired and input into the pre-defined radiotherapy instrument. This application achieves three-dimensional visualization of metabolic activity and immune response through multimodal fusion of spectral CT / functional magnetic resonance imaging and ResNet radiomics feature extraction. Compared to the identification accuracy of gross target volume (GTV), this method significantly improves accuracy. Furthermore, the immune thermogram generated by the XGBoost model and SAMed-2 segmentation markers can accurately locate highly responsive voxels. The Monte Carlo simulation of gradient dose distribution ensures that the dose intensity in the immune-activated region is several times that of the conventional region while protecting inactive regions. The multi-leaf grating motion trajectory optimization algorithm reduces the energy consumption of lead-gate drive in existing linear accelerators while maintaining compatibility with mainstream radiotherapy equipment. Moreover, the XGBoost model training can dynamically adjust the dose hotspot based on the patient's specific immune characteristics, ensuring that radiation accurately and sufficiently reaches the areas within the tumor region that most activate the immune system. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating the steps of an embodiment of the spatial segmentation radiotherapy guidance method based on target immune region activation according to this application.

[0075] Figure 2 This is a schematic diagram of the functional modules of a spatial segmentation radiotherapy guidance system based on target immune region activation according to an embodiment of this application;

[0076] Figure 3 This is a schematic diagram of the structure of an embodiment of the electronic device of this application;

[0077] Figure 4 This is a schematic diagram of the structure of one embodiment of the storage medium of this application. Detailed Implementation

[0078] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0079] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0080] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same instance, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0081] like Figure 1 As shown, this embodiment provides an example of a spatial segmentation radiotherapy guidance method based on target immune region activation. In this embodiment, the radiotherapy guidance method is applied to patients undergoing radiotherapy treatment using a pre-set radiotherapy instrument.

[0082] Preferably, the preset radiotherapy instrument can be upgraded in terms of software and hardware to directly implement the steps described below, or it can be connected to an external electronic device or computer to operate the preset radiotherapy instrument and complete the steps described below.

[0083] Preferably, for pre-installed radiotherapy instruments undergoing hardware and software upgrades, the following conditions must be met:

[0084] The hardware requirements are as follows:

[0085] ① Image acquisition system: It needs to be equipped with spectral CT (supporting multi-level imaging) and functional magnetic resonance imaging (such as DCE-MRI or DWI sequence), with spatial resolution ≤1mm³, dual-modal synchronous acquisition capability, and time registration error <0.5 seconds.

[0086] ② Treatment execution system: The linear accelerator must meet the following requirements: X-ray energy ≥6MV, maximum dose rate ≥330MU / min (supporting gradient dose output); multi-leaf grating (MLC) blade width ≤2.5mm, supporting dynamic intensity modulation; gantry rotation accuracy ≤0.5°, with 360° continuous rotation capability.

[0087] ③ Auxiliary equipment: translational accuracy of the six-dimensional treatment bed ≤0.5mm, rotational accuracy ≤0.1°; real-time image guidance system, such as CBCT or ultrasound positioning.

[0088] The software requirements are as follows:

[0089] ① Image processing module: The bilinear registration algorithm must support non-rigid deformation registration (such as the Elastix framework); the ResNet network requires CUDA 11.0 or higher and ≥16GB of video memory.

[0090] ② Treatment planning module: Monte Carlo dosing engine. For example, EGSnrc or GPUMCD; supports multi-target optimization, target coverage ≥95%, OAR dose ≤ tolerable dose; SAMed-2 model needs to integrate DICOM RT standard interface.

[0091] ③ Control and Management System: The motion control command sequence complies with the IEC 60601 safety standard; it has real-time dose verification functions, such as ArcCheck three-dimensional dose verification.

[0092] Specifically, this radiotherapy guidance method includes the following steps:

[0093] Step S1: Obtain the spectral CT and functional magnetic resonance imaging data of individual patients, and generate a fused image matrix dataset using a bilinear registration algorithm.

[0094] Preferably, the energy dispersive CT scanning parameters are dual-level scanning (80kVp / 140kVp), slice thickness ≤1mm; reconstruction matrix 512×512, FOV 350mm; matrix material decomposition map (water / iodine map) needs to be acquired simultaneously.

[0095] Preferably, the functional magnetic resonance imaging (fMRI) acquisition DWI sequence has a b-value of 0 / 800 / 1000 s / mm²; DCE-MRI has a temporal resolution of ≤5 s and a duration of 5 minutes; and a spatial resolution of 1×1×3 mm³.

[0096] The implementation process for generating the fused image matrix dataset using the bilinear registration algorithm is as follows:

[0097] ① Preprocessing stage:

[0098] Apply N4 bias field correction to the CT data.

[0099] Time-layer correction and head motion compensation were performed on the fMRI.

[0100] Unified resampling to 1 mm³ isotropic resolution.

[0101] ②The Python code for the registration core algorithm is as follows:

[0102] # Implementation of bilinear registration based on SimpleITK

[0103] import SimpleITK as sitk

[0104] def bilinear_registration(fixed_img, moving_img):

[0105] elastix = sitk.ElastixImageFilter()

[0106] elastix.SetFixedImage(fixed_img) # Sets the CT image as a fixed image

[0107] elastix.SetMovingImage(moving_img) # fMRI as a moving image

[0108] # Registration parameter settings

[0109] parameter_map = sitk.GetDefaultParameterMap("bspline")

[0110] parameter_map["Metric"] = ["AdvancedMattesMutualInformation"]

[0111] parameter_map["NumberOfResolutions"] = ["4"]

[0112] elastix.SetParameterMap(parameter_map)

[0113] elastix.Execute()

[0114] return elastix.GetResultImage()

[0115] ③ Fusion matrix generation:

[0116] A weighted fusion strategy was adopted: CT provided anatomical structure weight (70%), and fMRI provided functional information weight (30%).

[0117] Output matrix format: 4D tensor (x,y,z,channel) stored in HDF5.

[0118] Step S2: Extract the radiomics features of the fused image matrix dataset using a ResNet convolutional network.

[0119] Preferably, the architecture of the ResNet convolutional network is designed as follows:

[0120] ①ResNet network architecture design:

[0121] Basic module selection: ResNet18 structure (containing 17 convolutional layers + 1 fully connected layer).

[0122] Use BasicBlock residual blocks (for 3D medical imaging).

[0123] The number of input layer channels is set to the number of modalities of the fused image (e.g., 2 channels for CT+fMRI).

[0124] ② 3D convolution adaptation, Python code is as follows:

[0125] # 3D Residual Block Implementation (Based on PyTorch)

[0126] class Residual3D(nn.Module):

[0127] def __init__(self, in_channels, out_channels, stride=1):

[0128] super().__init__()

[0129] self.conv1 = nn.Conv3d(in_channels, out_channels, kernel_size=3,

[0130] stride=stride, padding=1)

[0131] self.bn1 = nn.BatchNorm3d(out_channels)

[0132] self.conv2 = nn.Conv3d(out_channels, out_channels, kernel_size=3,

[0133] padding=1)

[0134] self.bn2 = nn.BatchNorm3d(out_channels)

[0135] if stride != 1 or in_channels != out_channels:

[0136] self.shortcut = nn.Sequential(

[0137] nn.Conv3d(in_channels, out_channels, kernel_size=1,

[0138] stride=stride),

[0139] nn.BatchNorm3d(out_channels))

[0140] else:

[0141] self.shortcut = nn.Identity()

[0142] def forward(self, x):

[0143] residual = self.shortcut(x)

[0144] out = F.relu(self.bn1(self.conv1(x)))

[0145] out = self.bn2(self.conv2(out))

[0146] out += residual

[0147] return F.relu(out)

[0148] ③ Feature extraction process:

[0149] Preprocessing: Z-score normalization is performed on the fused images.

[0150] Network structure: Initial convolutional layer: 7×7×7 convolution, stride 2, output 64 channels; Max pooling: 3×3×3 kernel, stride 2; Residual block stacking: Layer 1: 2 residual blocks (64 channels), Layer 2: 2 residual blocks (128 channels, downsampled), Layer 3: 2 residual blocks (256 channels, downsampled), Layer 4: 2 residual blocks (512 channels, downsampled).

[0151] Global average pooling layer.

[0152] ④ Imageomics feature generation: Extract 256-dimensional depth feature vectors from Layer 3; use the mRMR algorithm to select the Top 50 features; reconstruct a 3D feature map through deconvolution, restoring the resolution to 1 / 8 of the input size.

[0153] Step S3: Identify metabolically active regions of radiomics features through Otsu adaptive threshold segmentation to obtain an immune activity feature vector set.

[0154] Preferably, the Otsu adaptive threshold segmentation algorithm (Otsu adaptive threshold algorithm) is implemented as follows:

[0155] ① 3D Image Adaptation: Voxel-level processing is performed on the 3D feature map (128×128×128) output by ResNet; a sliding window strategy (window size 16×16×16, stride 8) is used to achieve local adaptation. The Python code is as follows:

[0156] # 3D Otsu Implementation Based on SimpleITK

[0157] import SimpleITK as sitk

[0158] from skimage.filters import threshold_otsu

[0159] def otsu_3d(feature_volume):

[0160] otsu_filter = sitk.OtsuThresholdImageFilter()

[0161] otsu_filter.SetInsideValue(0)

[0162] otsu_filter.SetOutsideValue(1)

[0163] binary_volume = otsu_filter.Execute(feature_volume)

[0164] return sitk.GetArrayFromImage(binary_volume)

[0165] ② Identification of metabolically active regions: Calculate the inter-class variance σ for each sliding window. 2 =ω1ω2(μ1-μ2) 2

[0166] The window region where σ² > 0.3 is retained. It is worth noting that the formula for inter-class variance and the meaning of its symbols are generally known, and the meaning of the symbols will not be repeated in this embodiment.

[0167] ③ Feature vector generation: Extract connected regions (clusters with ≥50 voxels); calculate the 7-dimensional features of each region as feature vectors.

[0168] Preferably, the 7-dimensional features may include the following dimensions:

[0169] 1. Average gray value.

[0170] 2. Gray standard deviation.

[0171] 3. Volume (voxel count).

[0172] 4. Sphericity (36πV) 2 / S 3 ).

[0173] 5. Rate of change of surface curvature.

[0174] 6. Texture entropy.

[0175] 7. Neighborhood gradient mean.

[0176] Step S4: Train the XGBoost classification model using a preset dataset based on the XGBoost classification model, and input the immune activity feature vector set into the trained XGBoost classification model to obtain the immune heatmap.

[0177] Preferably, the XGBoost classification model is constructed as follows:

[0178] ① Basic parameter configuration, as shown in the Python code below:

[0179] import xgboost as xgb

[0180] params = {

[0181] 'objective': 'binary:logistic', # Binary classification task

[0182] 'eval_metric': 'auc',

[0183] 'max_depth': 6,

[0184] 'learning_rate': 0.01,

[0185] 'subsample': 0.8,

[0186] 'colsample_bytree': 0.7,

[0187] 'gamma': 0.1,

[0188] 'reg_alpha': 0.5, # L1 regular expression

[0189] 'reg_lambda': 1.0 # L2 regular expression

[0190] }

[0191] ② Feature engineering processing:

[0192] The immune activity feature vector is standardized (Z-score).

[0193] The SMOTE algorithm is used to solve the class imbalance problem.

[0194] Use mRMR features to select and retain the top 30 features.

[0195] Preferably, the training process for the XGBoost classification model is as follows:

[0196] The data preparation stage is shown in the Python code below:

[0197] # Load the preset dataset (must include immune activity annotations)

[0198] dtrain = xgb.DMatrix(X_train, label=y_train)

[0199] dtest = xgb.DMatrix(X_test, label=y_test)

[0200] # Early Stop Settings

[0201] watchlist = [(dtrain, 'train'), (dtest, 'eval')]

[0202] The model training uses 5-fold cross-validation and Bayesian optimization for hyperparameter search. The training termination condition can be set to no improvement in the validation set AUC for 10 consecutive rounds.

[0203] Preferably, the processing flow in the immune thermogram generation stage is as follows:

[0204] ① The predicted output processing is shown in the Python code below:

[0205] model = xgb.train(params, dtrain, num_boost_round=500,

[0206] evals=watchlist, early_stopping_rounds=10)

[0207] # Obtaining voxel-level prediction probabilities

[0208] prob_map = model.predict(xgb.DMatrix(feature_vectors))

[0209] ② Visualization of immune thermograms:

[0210] Map the probability values ​​to the gray range of [0, 255].

[0211] Use bilinear interpolation to restore the original image resolution.

[0212] Superimposed pseudo-color coding (red-yellow gradient indicates the intensity of immune activity).

[0213] Step S5: Obtain the radiotherapy localization CT of the patient and align it with the immune thermogram coordinate system. Then, use the SAMed-2 image segmentation model to perform voxel-level segmentation on the aligned immune thermogram and mark the voxels with high immune response.

[0214] Preferably, the specific implementation method of coordinate system alignment is as follows:

[0215] ① Spatial registration technology:

[0216] A rigid registration algorithm based on mutual information is adopted.

[0217] The following Python code demonstrates how to implement DICOM RT coordinate system transformation using the ITK library:

[0218] import SimpleITK as sitk

[0219] def align_coordinates(ct_volume, heatmap):

[0220] # Initialize the registration unit

[0221] registration = sitk.ImageRegistrationMethod()

[0222] registration.SetMetricAsMattesMutualInformation()

[0223] registration.SetOptimizerAsRegularStepGradientDescent()

[0224] # Perform registration

[0225] transform = registration.Execute(ct_volume, heatmap)

[0226] aligned_heatmap = sitk.Resample(heatmap, ct_volume, transform)

[0227] return aligned_heatmap

[0228] Preferably, the SAMed-2 image segmentation model performs voxel-level segmentation on the aligned immune heatmap as follows:

[0229] ① Model loading and configuration:

[0230] Use a pre-trained version of SAMed-2-Base (input size 256×256×256).

[0231] The key parameter settings are as follows:

[0232] Batch size: 4

[0233] Learning rate: 2e-5

[0234] Number of iterations: 1000

[0235] Loss function: Dice + Focal Loss

[0236] ② Voxel-level segmentation process:

[0237] During the input preprocessing stage, the heatmap is normalized to the [0,1] interval, and Gaussian noise (σ=0.01) is added to enhance robustness.

[0238] The multi-scale inference strategy employs a joint prediction using the original resolution plus a 2x downsampled version, and fuses the results through a voting mechanism.

[0239] Preferably, the specific implementation method of immune high-response voxel labeling is as follows:

[0240] ①Feature extraction criteria must simultaneously meet the following conditions:

[0241] Heatmap value ≥ 0.7 (after normalization).

[0242] The volume of the connected region is ≥8mm³.

[0243] Gradient change rate > 15% / mm.

[0244] ② The output format specification requires the generation of a DICOM RT Structure file to ensure compatibility.

[0245] Example as follows:

[0246] (3006,2020) ROI Number

[0247] (3006,2022) ROI Name: "High Immune Response Zone"

[0248] (3006,2040) ROI Volume

[0249] Step S6: Obtain the three-dimensional coordinates of all highly responsive immune voxels and integrate them into a three-dimensional coordinate set of the immune target area. Iterate the motion trajectory of the multi-leaf grating of the preset radiotherapy instrument through Monte Carlo simulation, and cover all three-dimensional coordinates with the rays of the preset radiotherapy instrument in the form of a preset ray intensity gradient decreasing.

[0250] Preferably, the three-dimensional coordinate integration of the immune target region is implemented as follows:

[0251] ① Coordinate system transformation:

[0252] Convert the voxel coordinates output by SAMed-2 to the DICOM RT coordinate system (unit: mm).

[0253] The Python code for establishing the mapping relationship between voxel indices and physical coordinates is as follows:

[0254] def voxel_to_world(voxel_coord, origin, spacing):

[0255] return [o + v*s for o,v,s in zip(origin, voxel_coord, spacing)]

[0256] ② Target region clustering processing:

[0257] Outliers were removed using the DBSCAN algorithm (eps=5mm, min_samples=10).

[0258] The centroid coordinates of each cluster are calculated as key control points.

[0259] Preferably, the Monte Carlo simulation is implemented as follows:

[0260] ① Multileaf grating modeling:

[0261] The MLC model was constructed based on the parameters of the VARIAN RapidArc system: 120 blades (5mm width in the central area), maximum movement speed of 4cm / s, and positioning accuracy of ±0.5mm.

[0262] ② Dose distribution simulation, as shown in the Python code below:

[0263] import monaco

[0264] def mc_simulation(target_coords):

[0265] # Initialize the simulation environment

[0266] sim = monaco.MCSimulation(particles=1e6)

[0267] # Set X-ray source parameters (6MV-X-ray)

[0268] sim.set_beam(energy=6, ssd=100, field_size=(40,40))

[0269] # Define intensity gradient: 100% at the center → 20% at the edge

[0270] dose_profile = lambda r: 1 - 0.8*(r / max_radius)

[0271] # Perform iterative calculations

[0272] results = sim.run(target_coords, dose_profile)

[0273] return results.optimized_trajectory

[0274] Preferably, the motion trajectory of the multi-leaf grating is optimized using Pareto multi-objective optimization, with objective 1 set as target area coverage ≥ 95% and objective 2 as OAR dose ≤ tolerable dose, while defining the constraint condition as leaf acceleration < 3 cm / s². 2 .

[0275] Preferably, the control commands for the motion trajectory of the multi-leaf grating are generated using a DICOM RT Plan format command sequence, as shown in the example below:

[0276] CONTROL POINT 1:

[0277] Gantry Angle: 0.0°

[0278] MLC Positions: [12.3,14.5,...,8.7] (mm)

[0279] Dose Rate: 300 MU / min

[0280] Step S7: Obtain the motion control command sequence after the multi-leaf grating motion trajectory iteration is completed and input it into the preset radiotherapy instrument.

[0281] Preferably, the trajectory data output by Monte Carlo can be converted into the standard DICOM RT Plan structure, as shown in the following Python code:

[0282] def generate_rtplan(trajectory):

[0283] plan = pydicom.Dataset()

[0284] plan.SOPClassUID = '*************' # RT Plan

[0285] plan.ControlPointSequence = [

[0286] create_control_point(cp) for cp in trajectory ]

[0288] return plan

[0289] Each control point includes the rack angle (0.1° accuracy), MLC blade position (0.5mm accuracy), and dose rate (MU / min).

[0290] Further, in step S7, the motion control command sequence after the multi-leaf grating motion trajectory iteration is completed is obtained and input into the preset radiotherapy instrument. Afterwards, the following steps are also included:

[0291] Step S10: After the radiotherapy treatment is completed by the preset radiotherapy instrument, obtain the SUVmax value matrix of the individual patient based on FDG-PET / CT scan.

[0292] Preferably, FDG-PET / CT requires intravenous injection of 18F-FDG tracer. The specific scanning procedure for FDG-PET / CT is a mature existing technology, and this embodiment will not elaborate on the scanning procedure.

[0293] Preferably, the generation of the SUVmax matrix requires the following steps:

[0294] ① Data preprocessing, Python code as follows:

[0295] import pydicom

[0296] def load_petct(series_path):

[0297] # Read DICOM series data

[0298] series = [pydicom.dcmread(f) for f in series_path]

[0299] pet_data = np.stack([s.pixel_array * s.RescaleSlope +s.RescaleIntercept

[0300] for s in series if s.Modality == 'PT'])

[0301] ct_data = np.stack([s.pixel_array for s in series if s.Modality == 'CT'])

[0302] return pet_data, ct_data

[0303] ② Calculation of standardized uptake value: SUV = [tissue radioactivity concentration (kBq / ml)] / [injection dose (MBq) / body weight (kg)].

[0304] The matrix generation process includes attenuation correction (based on CT data), time attenuation correction (considering the half-life of radionuclides), and conversion of voxel values ​​to SUVs.

[0305] ③ Target region matrix extraction:

[0306] Registration with radiotherapy localization CT (mutual information method, accuracy ±1mm).

[0307] The Python code for extracting target voxels using a ROI mask is as follows:

[0308] def extract_target_suv(pet_array, mask_array):

[0309] return pet_array[mask_array > 0] # mask_array is the immune target area output by S5.

[0310] Step S20: Calculate the target area FDG uptake change rate based on the SUVmax value matrix using the baseline SUV.

[0311] Preferably, the baseline SUV needs to be acquired by the first FDG-PET / CT scan within 72 hours prior to treatment.

[0312] It is worth noting that the same equipment model, scanning protocol, and tracer injection dose must be used for FDG-PET / CT scans, both before and after treatment.

[0313] Preferably, the target area FDG uptake change rate is non-rigidly registered using the Elastix toolbox. Since the target area FDG uptake change rate is based on the SUVmax value matrix, the target area FDG uptake change rate can also be expressed in matrix form.

[0314] Step S30: Determine whether the change rate of FDG uptake in the target area exceeds the preset change rate threshold. If not, proceed to step S40.

[0315] Preferably, the preset threshold for the rate of change of FDG uptake in the target area can be set to 25%.

[0316] Step S40: Determine that the motion control command sequence is invalid radiotherapy guidance.

[0317] In step S50, in response to ineffective radiotherapy guidance, steps S4 to S6 are repeated using an additional preset sample set as the execution subject to continue iteratively updating the XGBoost classification model.

[0318] Preferably, invalid radiotherapy guidance may indicate a problem with the sample, so this embodiment introduces a new sample, which is maintained using a sliding window mechanism. The Python code is as follows:

[0319] class DynamicDataset:

[0320] def __init__(self, max_samples=1000):

[0321] self.queue = deque(maxlen=max_samples)

[0322] def add_batch(self, new_data):

[0323] # New data format: [features, label, patient_id]

[0324] self.queue.extend(new_data)

[0325] At the same time, new samples are assigned a weight of 1.5 times, while the weight of previously used historical samples decays over time.

[0326] Preferably, deep features can be extracted based on the FM-LCT pre-trained model to enhance feature engineering. The Python code is as follows:

[0327] from transformers import AutoModel

[0328] feature_extractor = AutoModel.from_pretrained("FM-LCT")

[0329] Preferably, the following strategy can be adopted to continue iteratively updating the XGBoost classification model:

[0330] ① Parameter optimization strategy:

[0331] The Bayesian hyperparameter search space, Python code is as follows:

[0332] param_space = {

[0333] 'max_depth': (3, 10),

[0334] 'learning_rate': (0.01, 0.3),

[0335] 'subsample': (0.6, 1.0)

[0336] }

[0337] ② Model update protocol, Python code as follows:

[0338] def incremental_train(old_model, new_data):

[0339] # Inheriting existing model structure

[0340] new_model = xgb.XGBClassifier(**old_model.get_params())

[0341] # Incremental Training

[0342] new_model.fit(

[0343] new_data['X'],

[0344] new_data['y'],

[0345] xgb_model=old_model )

[0347] return new_model

[0348] Step S60: Obtain the motion control command sequence after further iteration and input it into the preset radiotherapy instrument.

[0349] Step S70: Repeat steps S10 to S60 until the change rate of FDG uptake in the target area exceeds the preset change rate threshold.

[0350] Further, step S2, extracting radiomics features from the fused image matrix dataset using a ResNet convolutional network, specifically includes the following steps:

[0351] Step S21: Normalize the fused image matrix dataset.

[0352] Preferably, Z-score normalization can be used.

[0353] Step S22: Divide the normalized fused image matrix dataset into cubes of a preset size.

[0354] Preferably, the fused image matrix (512×512×64) is normalized (μ=0, σ=1) and divided into 64×64×64 cubic patches. N4 bias field correction is used to eliminate MRI artifacts, and the CT value cutoff range is [-1000, 2000] HU.

[0355] Step S23: Define the ResNet convolutional network and define the number of channels in the input layer of the ResNet convolutional network to match the preset size.

[0356] Preferably, the ResNet convolutional network is defined as shown in the following Python code:

[0357] # Add a self-attention layer to the ResNet infrastructure

[0358] class SA_ResNet(nn.Module):

[0359] def __init__(self):

[0360] super().__init__()

[0361] self.resnet = resnet50(pretrained=True)

[0362] self.attention = SelfAttention(2048) # Insert before the last layer

[0363] def forward(self, x):

[0364] x = self.resnet.conv1(x)

[0365] x = self.resnet.bn1(x)

[0366] x = self.resnet.relu(x)

[0367] x = self.resnet.maxpool(x)

[0368] x = self.resnet.layer1(x)

[0369] x = self.resnet.layer2(x)

[0370] x = self.resnet.layer3(x)

[0371] x = self.resnet.layer4(x)

[0372] x = self.attention(x) # Self-attention enhancement feature

[0373] return x

[0374] The input channel count is 64, matching the depth of the fusion matrix; a self-attention mechanism layer is inserted after layer 4.

[0375] Step S24: Input the fused image matrix dataset into the input layer of the ResNet convolutional network, and obtain a multidimensional deep feature vector set through forward propagation of the ResNet convolutional network.

[0376] Preferably, forward propagation obtains a 2048-dimensional deep feature vector, which is then reduced to 512 dimensions using GAP (Global Average Pooling).

[0377] Preferably, three sets of key features can be extracted:

[0378] First-order statistical characteristics (mean, variance, skewness, etc., 18 in total).

[0379] Texture features (22 features including contrast and correlation of GLCM).

[0380] Shape characteristics (8 in total, including volume and surface area).

[0381] Step S25: Filter the top feature vectors in the multidimensional deep feature vector set using the mRMR algorithm.

[0382] Preferably, the top 100 features are selected by mRMR (minimum redundancy maximum correlation) algorithm, and the dimensionality is reduced to 50-dimensional principal components (cumulative contribution rate > 95%) by PCA.

[0383] Step S26: All the top feature vectors are upsampled through a deconvolution layer to reconstruct a three-dimensional feature map, and the three-dimensional feature map is defined as an image omics feature.

[0384] Preferably, the design intent of steps S21 to S26 in this embodiment is to adopt an improved SA-ResNet architecture, adding a self-attention module on the basis of the traditional ResNet-50 to improve the ability to identify small lesions. The feature extraction process includes 1098 radiomics parameters, covering all categories defined by the International Radiomics Standard (IBSI). Furthermore, this embodiment can use the TCIA dataset (10,000 tumor images) published by NIH for pre-training through a transfer learning strategy, without the need to prepare a training dataset independently.

[0385] Further, step S3 involves identifying metabolically active regions of radiomics features through Otsu adaptive threshold segmentation to obtain an immune activity feature vector set, specifically including the following steps:

[0386] Step S31: Normalize the grayscale of the three-dimensional feature map and calculate the histogram distribution of the three-dimensional feature map after grayscale normalization.

[0387] Preferably, the image omics feature map (512×512×64) extracted by ResNet is grayscale normalized (0-255), the histogram distribution of the feature map is calculated, and the bimodal characteristic is verified to be a peak-to-valley ratio > 2:1.

[0388] Step S32: Calculate the inter-class variance of the histogram distribution using the Otsu adaptive threshold.

[0389] Preferably, the specific implementation for calculating the inter-class variance of the histogram distribution is shown in the following Python code:

[0390] def otsu_threshold(image):

[0391] total_pixels = image.size

[0392] hist = cv2.calcHist([image],[0],None,

[256] ,[0,256])

[0393] max_var = 0

[0394] optimal_thresh = 0

[0395] for t in range(1,255):

[0396] # Calculate within-class probability

[0397] w0 = np.sum(hist[:t]) / total_pixels

[0398] w1 = 1 - w0

[0399] # Calculate the class mean

[0400] if w0 > 0: mu0 = np.sum(np.arange(t) * hist[:t]) / (w0 *total_pixels)

[0401] else: mu0 = 0

[0402] if w1 > 0: mu1 = np.sum(np.arange(t,256) * hist[t:]) / (w1 *total_pixels)

[0403] else: mu1 = 0

[0404] # Calculate inter-class variance

[0405] var = w0 * w1 * (mu0 - mu1)**2

[0406] if var > max_var:

[0407] max_var = var

[0408] optimal_thresh = t

[0409] return optimal_thresh

[0410] Step S33: Select the optimal threshold among all candidate thresholds that maximizes the inter-class variance, and segment the three-dimensional feature map according to the optimal threshold to form high-metabolic regions and low-metabolic regions.

[0411] Preferably, Otsu threshold segmentation is applied to obtain a binary mask (1 = metabolically active region, 0 = metabolically inactive region), and micro-holes can be eliminated by morphological closing operation (3×3 circular kernel).

[0412] Step S34: Define the high-metabolic region as the metabolically active region.

[0413] Step S35: Based on the preset structural voxels, perform morphological closing operations on the metabolically active regions to obtain several connected regions.

[0414] Preferably, after eliminating micro-voids, the infiltration index of each connected region can be calculated using a formula:

[0415] ,in, For the first Infiltration index of each connected region The maximum standard intake value, For the effective metabolic biological volume, 1.2 is an exponential weighting. The mean value of the tumor background (or the mean value of the metabolically inactive area) is used to screen areas with an infiltration index > 0.7 as high-invasive areas.

[0416] Step S36: Calculate the gray-level co-occurrence matrix features for each connected region.

[0417] Preferably, the gray-level co-occurrence matrix features can be represented by the following three types of key parameters:

[0418] 1. Spatial characteristics: volume ratio of high infiltration zone, centroid coordinate variance.

[0419] 2. Metabolic characteristics: SUVmax / mean, total glycolysis (TLG).

[0420] 3. Heterogeneity features: contrast + correlation of the gray-level co-occurrence matrix (GLCM).

[0421] Step S37: Convert all gray-level co-occurrence matrix features into multidimensional feature vectors using a preset Python script.

[0422] Preferably, the preset Python script is as follows:

[0423] # Example Feature Extraction Core Code

[0424] import pyradiomics

[0425] from skimage.filters import threshold_otsu

[0426] def extract_immune_features(feature_volume):

[0427] # Otsu Multithreshold Segmentation

[0428] thresholds = threshold_multiotsu(feature_volume)

[0429] regions = np.digitize(feature_volume, bins=thresholds)

[0430] # Radiomics Feature Extraction

[0431] extractor = pyradiomics.featureextractor.RadiomicsFeatureExtractor()

[0432] results = extractor.execute(feature_volume, regions)

[0433] # Constructing feature vectors

[0434] immune_features = [

[0435] results['original_glcm_Contrast'],

[0436] results['original_glszm_ZoneVariance'],

[0437] calculate_metabolic_gradient(feature_volume), # Custom metabolic gradient calculation

[0438] regions.max() - regions.min() # Metabolically active range ]

[0440] return np.array(immune_features)

[0441] Step S38: Integrate all multidimensional feature vectors and define them as the immune activity feature vector set.

[0442] Preferably, a 50-dimensional immune activity feature vector can be generated through Z-score normalization.

[0443] Further, in step S4, the XGBoost classification model is trained using a pre-set dataset, and the immune activity feature vector set is input into the trained XGBoost classification model to obtain the immune heatmap. This specifically includes the following steps:

[0444] Step S41: Define the model parameters of the XGBoost classification model based on the preset model strategy.

[0445] Preferably, the model parameters of the XGBoost classification model can be defined as:

[0446] The base learning rate is set to 0.01.

[0447] The maximum tree depth is 6 levels.

[0448] gbtree is used as the booster type.

[0449] Loss function: Use binary cross-entropy loss.

[0450] Enable early stopping mechanism: terminate training when the validation set AUC does not improve for 5 consecutive rounds.

[0451] Step S42: Standardize and reduce the dimensions of the preset dataset, and input the sample sets in the preset dataset into the XGBoost classification model in batches for training.

[0452] Step S43: In each batch of training, the XGBoost classification model is validated using the validation set in the preset dataset, and training is terminated when the AUC value of the validation set no longer increases, thus obtaining the trained XGBoost classification model.

[0453] Preferably, the training process of the XGBoost classification model is described in the relevant explanation of step S4 above, and will not be repeated here.

[0454] Step S44: Input the immune activity feature vector set into the trained XGBoost classification model and obtain the positive probability prediction value for each voxel.

[0455] Preferably, the immune activity feature vector of the case to be predicted is input into the trained XGBoost model to obtain the positive probability prediction value of each voxel. The positive probability prediction value ranges from [0,1], and the probability value is mapped back to the three-dimensional matrix of the original image coordinate system.

[0456] Step S45: Calculate the numerical range of all positive probability prediction values ​​and assign the numerical range to at least three consecutive and different heatmap color levels.

[0457] Preferably, the heatmap color scale can be defined as: blue (below 0.3) - yellow (0.3-0.7) - red (above 0.7).

[0458] Step S46: Map all positive probability prediction values ​​and corresponding heatmap color levels to the coordinate system three-dimensional matrix of the fused image matrix dataset to obtain the immune heatmap.

[0459] Preferably, bilinear interpolation can be used to smooth the probability boundaries to obtain a DICOM format heatmap file.

[0460] Further, in step S5, the patient's radiotherapy localization CT scan is acquired and aligned with the immune thermogram coordinate system. The aligned immune thermogram is then segmented at the voxel level using the SAMed-2 image segmentation model, while marking highly immune-responsive voxels. This specifically includes the following steps:

[0461] Step S51: Obtain the coordinate system centroid of the radiotherapy localization CT and the coordinate system centroid of the immunothermal map.

[0462] Step S52: Calculate the rotation and translation matrix of the radiotherapy positioning CT and the immunothermograph based on the centroids of the two coordinate systems.

[0463] Preferably, the calculation process for the rotation and translation matrix in the dual coordinate system is as follows:

[0464] ① Calculate the centroids of the point sets in the two coordinate systems respectively.

[0465] ②Decentralize the two point sets.

[0466] ③ Construct a covariance matrix that describes the spatial relationship between two point sets.

[0467] ④ Perform singular value decomposition (SVD) on the covariance matrix.

[0468] ⑤ Calculate the translation vector.

[0469] It is worth noting that the rotation and translation matrix is ​​a common calculation method between coordinate systems, and the specific calculation formula of the calculation process will not be described in this embodiment.

[0470] Step S53: Based on the rotation and translation matrix, the radiotherapy localization CT and immunothermal map are registered using bicubic interpolation to achieve coordinate system alignment.

[0471] Preferably, for non-integer coordinate positions, interpolation calculations are performed using at least 26 surrounding voxels, with interpolation weights based on cubic polynomial functions, and cubic interpolation calculations are performed in the x, y, and z directions respectively to map the heatmap to the CT coordinate system.

[0472] Step S54: The aligned immune heatmap is segmented at the voxel level using the SAMed-2 image segmentation model.

[0473] Preferably, the voxel-level segmentation of the SAMed-2 image segmentation model is described above.

[0474] Step S55: Mark all voxels in the high-metabolic region as highly immune-responsive voxels.

[0475] Further, in step S6, the three-dimensional coordinates of all highly responsive immune voxels are acquired and integrated into a three-dimensional coordinate set of the immune target area. The motion trajectory of the preset radiotherapy instrument's multi-leaf grating is iteratively simulated using Monte Carlo simulation, and the radiation from the preset radiotherapy instrument is used to cover all three-dimensional coordinates in a preset radiation intensity gradient decreasing manner, including:

[0476] Step S61: Extract the three-dimensional coordinates of all highly responsive immune voxels from the immune heatmap segmented by the SAMed-2 image segmentation model.

[0477] Step S62: Generate the boundary surface based on all three-dimensional coordinates using the Delaunay triangulation algorithm.

[0478] Step S63: Using the center of the boundary surface as the reference point, set the dose attenuation coefficient according to the preset interval distance.

[0479] Preferably, the dose gradient parameter is set as follows: with the center of the target area as the reference, the dose attenuation coefficient is set at 5mm intervals (0.8→0.6→0.4).

[0480] Step S64: Determine the initial opening and closing state of the multi-leaf grating and the initial angle of the collimator of the radiotherapy instrument based on the projection profile of the boundary surface.

[0481] Step S65: Using the initial angle of the collimator as a reference, the optimal opening and closing combination of the multi-leaf grating at each angle of the collimator is calculated iteratively using a global optimization algorithm with a preset angle step size. All optimal opening and closing combinations are based on the projection profile that can just completely cover the boundary surface.

[0482] Step S66: Using the motion time sequence of the collimator as the chronological order, convert the position sequence of all optimal opening and closing combinations of the multileaf grating into the motion trajectory of the multileaf grating.

[0483] Preferably, the global optimization algorithm can be set as the ant colony algorithm.

[0484] This embodiment acquires spectral CT and functional magnetic resonance imaging (fMRI) data of individual patients and generates a fused image matrix dataset using a bilinear registration algorithm. Radiomics features of the fused image matrix dataset are extracted using a ResNet convolutional network. Metabolicly active regions of the radiomics features are identified using Otsu adaptive threshold segmentation to obtain an immune activity feature vector set. An XGBoost classification model is trained using a pre-defined dataset, and the immune activity feature vector set is input into the trained XGBoost model to obtain an immune heatmap. The patient's radiotherapy localization CT scan is acquired and aligned with the immune heatmap coordinate system. The aligned immune heatmap is then segmented at the voxel level using a SAMed-2 image segmentation model, while marking highly responsive voxels. The three-dimensional coordinates of all highly responsive voxels are acquired and integrated into a three-dimensional coordinate set of the immune target area. The motion trajectory of the pre-defined radiotherapy instrument is iterated using Monte Carlo simulation, ensuring that the radiation from the pre-defined radiotherapy instrument covers all three-dimensional coordinates in a decreasing gradient manner. The motion control command sequence after the multi-leaf grating motion trajectory iteration is completed is acquired and input into the pre-defined radiotherapy instrument. This embodiment achieves three-dimensional visualization of metabolic activity and immune response through multimodal fusion of spectral CT / functional magnetic resonance imaging and ResNet radiomics feature extraction. Compared with the identification accuracy of anatomical target volume (GTV), it has a significant improvement. In addition, the immune thermogram generated by the XGBoost model and the SAMed-2 segmentation markers can accurately locate the immune high-response voxels. The gradient dose distribution simulated by Monte Carlo simulation makes the dose intensity of the immune activation area several times that of the conventional area while protecting the inactive area. The multi-leaf grating motion trajectory optimization algorithm of this embodiment can reduce the energy consumption of the lead gate drive of the existing linear accelerator while being compatible with mainstream radiotherapy equipment. Moreover, the XGBoost model training can dynamically adjust the dose hotspot according to the patient's specific immune characteristics, so that the radiation can accurately and sufficiently reach the area in the tumor region that is most likely to activate the immune system.

[0485] like Figure 2 As shown, this embodiment provides a spatial segmentation radiotherapy guidance system based on target immune region activation, which is applied to the radiotherapy guidance method as described in the above method embodiment.

[0486] Specifically, the radiotherapy guidance system includes a fusion image matrix dataset acquisition module 1, a radiomics feature extraction module 2, an immune activity feature vector set acquisition module 3, an immune thermogram acquisition module 4, an immune high-response voxel labeling module 5, a multi-leaf grating motion trajectory iteration module 6, and a motion control command sequence output module 7.

[0487] The system comprises the following modules: Module 1 for acquiring spectral CT and functional magnetic resonance imaging (fMRI) data of individual patients, and generating a fused image matrix dataset using a bilinear registration algorithm; Module 2 for extracting radiomics features from the fused image matrix dataset using a ResNet convolutional network; Module 3 for acquiring immune activity feature vector sets, identifying metabolically active regions of radiomics features using Otsu adaptive threshold segmentation to obtain an immune activity feature vector set; and Module 4 for acquiring immune heatmaps, training an XGBoost classification model based on a preset dataset, and inputting the immune activity feature vector set into the trained XGBoost classification model to obtain the immune heatmap. The module 5, which is responsible for the thermal mapping, is used to acquire the radiotherapy positioning CT of the patient and align it with the coordinate system of the immunothermal map. The aligned immunothermal map is then segmented at the voxel level using the SAMed-2 image segmentation model, and the immunoreactive voxels are marked. The module 6, which is responsible for the multi-leaf grating motion trajectory iteration, is used to acquire the three-dimensional coordinates of all immunoreactive voxels and integrate them into a three-dimensional coordinate set of the immune target area. The module iterates the motion trajectory of the preset radiotherapy instrument using Monte Carlo simulation, and covers all three-dimensional coordinates with the rays of the preset radiotherapy instrument in a preset ray intensity gradient decreasing form. The module 7, which is responsible for the motion control command sequence output, is used to acquire the motion control command sequence after the multi-leaf grating motion trajectory iteration is completed and input it into the preset radiotherapy instrument.

[0488] Furthermore, the radiotherapy guidance system also includes, in sequence, an SUVmax value matrix acquisition module, a target area FDG uptake change rate calculation module, a target area FDG uptake change rate judgment module, an ineffective radiotherapy guidance judgment module, an XGBoost classification model re-iteration module, an iterated motion control command sequence input module, and a repeated iteration module; the SUVmax value matrix acquisition module and the motion control command sequence output module 7 are electrically or signal connected.

[0489] The system includes the following modules: SUVmax value matrix acquisition module, which acquires the SUVmax value matrix of an individual patient based on FDG-PET / CT scans after radiotherapy treatment with a preset radiotherapy instrument; target area FDG uptake change rate calculation module, which calculates the target area FDG uptake change rate based on the baseline SUV value matrix; target area FDG uptake change rate judgment module, which determines whether the target area FDG uptake change rate exceeds a preset change rate threshold; invalid radiotherapy guidance judgment module, which determines the motion control command sequence as invalid radiotherapy guidance if not; XGBoost classification model re-iteration module, which, in response to invalid radiotherapy guidance, repeatedly executes the immunothermal map acquisition module 4 to the multi-leaf grating motion trajectory iteration module 6 using an additional preset sample set as the execution subject to continue iterating and updating the XGBoost classification model; post-iteration motion control command sequence input module, which acquires the motion control command sequence after further iteration and inputs it into the preset radiotherapy instrument; and repeated iteration module, which repeatedly executes the SUVmax value matrix acquisition module to the post-iteration motion control command sequence input module until the target area FDG uptake change rate exceeds the preset change rate threshold.

[0490] Furthermore, the radiomics feature extraction module 2 specifically includes a first radiomics feature extraction unit, a second radiomics feature extraction unit, a third radiomics feature extraction unit, a fourth radiomics feature extraction unit, a fifth radiomics feature extraction unit, and a sixth radiomics feature extraction unit that are electrically or signal-connected in sequence; the first radiomics feature extraction unit is electrically or signal-connected to the fused image matrix dataset acquisition module 1, and the sixth radiomics feature extraction unit is electrically or signal-connected to the immune activity feature vector set acquisition module 3.

[0491] The system comprises the following components: a first radiomics feature extraction unit, used to normalize the fused image matrix dataset; a second radiomics feature extraction unit, used to segment the normalized fused image matrix dataset into cubes of a preset size; a third radiomics feature extraction unit, used to define a ResNet convolutional network and ensure that the number of channels in the input layer of the ResNet convolutional network matches the preset size; a fourth radiomics feature extraction unit, used to input the fused image matrix dataset into the input layer of the ResNet convolutional network and obtain a multidimensional deep feature vector set through forward propagation of the ResNet convolutional network; a fifth radiomics feature extraction unit, used to filter the top feature vectors in the multidimensional deep feature vector set using the mRMR algorithm; and a sixth radiomics feature extraction unit, used to upsample all the top feature vectors through a deconvolutional layer to reconstruct a three-dimensional feature map and define the three-dimensional feature map as a radiomics feature.

[0492] Furthermore, the immune activity feature vector set acquisition module 3 specifically includes a first immune activity feature vector set acquisition unit, a second immune activity feature vector set acquisition unit, a third immune activity feature vector set acquisition unit, a fourth immune activity feature vector set acquisition unit, a fifth immune activity feature vector set acquisition unit, a sixth immune activity feature vector set acquisition unit, a seventh immune activity feature vector set acquisition unit, and an eighth immune activity feature vector set acquisition unit that are electrically or signal-connected in sequence. The first immune activity feature vector set acquisition unit is electrically or signal-connected to the sixth radiomics feature extraction unit, and the eighth immune activity feature vector set acquisition unit is electrically or signal-connected to the immune thermogram acquisition module 4.

[0493] The system comprises the following components: a first immune activity feature vector set acquisition unit, a second immune activity feature vector set acquisition unit, and a third immune activity feature vector set acquisition unit. The first unit performs gray-level normalization on the 3D feature map and calculates the histogram distribution of the normalized 3D feature map. The second unit calculates the inter-class variance of the histogram distribution using an Otsu adaptive threshold. The third unit selects the optimal threshold from all candidate thresholds that maximizes the inter-class variance and segments the 3D feature map according to the optimal threshold, forming high-metabolic and low-metabolic regions. The fourth unit defines the high-metabolic region as a metabolically active region. The fifth unit performs morphological closing operations on the metabolically active region based on a preset structure voxel to obtain several connected regions. The sixth unit calculates the gray-level co-occurrence matrix features of each connected region. The seventh unit converts all gray-level co-occurrence matrix features into multi-dimensional feature vectors using a preset Python script. The eighth unit integrates all multi-dimensional feature vectors and defines them as an immune activity feature vector set.

[0494] Furthermore, the immune thermogram acquisition module 4 specifically includes a first immune thermogram acquisition unit, a second immune thermogram acquisition unit, a third immune thermogram acquisition unit, a fourth immune thermogram acquisition unit, a fifth immune thermogram acquisition unit, and a sixth immune thermogram acquisition unit that are electrically or signal-connected in sequence; the first immune thermogram acquisition unit is electrically or signal-connected to the eighth immune activity feature vector set acquisition unit, and the sixth immune thermogram acquisition unit is electrically or signal-connected to the immune high-response voxel labeling module 5.

[0495] The system comprises the following components: a first immune heatmap acquisition unit, used to define the model parameters of the XGBoost classification model based on a preset model strategy; a second immune heatmap acquisition unit, used to standardize and reduce the dimensions of the preset dataset, and input the sample set from the preset dataset into the XGBoost classification model in batches for training; a third immune heatmap acquisition unit, used to validate the XGBoost classification model using the validation set in the preset dataset in each batch of training, and to terminate training when the AUC value of the validation set no longer increases, thus obtaining the trained XGBoost classification model; a fourth immune heatmap acquisition unit, used to input the immune activity feature vector set into the trained XGBoost classification model, and to obtain the positive probability prediction value for each voxel; a fifth immune heatmap acquisition unit, used to statistically analyze the numerical range of all positive probability prediction values, and to assign at least three consecutive and different heatmap color levels to the numerical range; and a sixth immune heatmap acquisition unit, used to map all positive probability prediction values ​​and their corresponding heatmap color levels to the coordinate system three-dimensional matrix of the fused image matrix dataset, thus obtaining the immune heatmap.

[0496] Furthermore, the immune high-response voxel labeling module 5 specifically includes a first immune high-response voxel labeling unit, a second immune high-response voxel labeling unit, a third immune high-response voxel labeling unit, a fourth immune high-response voxel labeling unit, and a fifth immune high-response voxel labeling unit that are electrically connected in sequence; the first immune high-response voxel labeling unit is electrically or signal-connected to the sixth immune thermogram acquisition unit, and the fifth immune high-response voxel labeling unit is electrically or signal-connected to the multi-leaf grating motion trajectory iteration module 6.

[0497] The system comprises the following components: a first high-response voxel labeling unit for obtaining the coordinate centroids of the radiotherapy localization CT and the immunothermograph; a second high-response voxel labeling unit for calculating the rotation and translation matrices of the radiotherapy localization CT and the immunothermograph based on the two coordinate centroids; a third high-response voxel labeling unit for registering the radiotherapy localization CT and the immunothermograph using bicubic interpolation based on the rotation and translation matrices to achieve coordinate system alignment; a fourth high-response voxel labeling unit for voxel-level segmentation of the aligned immunothermograph using the SAMed-2 image segmentation model; and a fifth high-response voxel labeling unit for labeling all voxels in high-metabolic regions as high-response voxels.

[0498] Furthermore, the multi-leaf grating motion trajectory iteration module 6 specifically includes a first multi-leaf grating motion trajectory iteration unit, a second multi-leaf grating motion trajectory iteration unit, a third multi-leaf grating motion trajectory iteration unit, a fourth multi-leaf grating motion trajectory iteration unit, a fifth multi-leaf grating motion trajectory iteration unit, and a sixth multi-leaf grating motion trajectory iteration unit that are electrically or signal-connected in sequence; the first multi-leaf grating motion trajectory iteration unit is electrically or signal-connected to the fifth immune high-response voxel labeling unit, and the sixth multi-leaf grating motion trajectory iteration unit is electrically or signal-connected to the motion control command sequence output module 7.

[0499] The system comprises the following components: a first multileaf grating motion trajectory iteration unit, used to extract the three-dimensional coordinates of all highly responsive immune voxels from the immunothermal map segmented by the SAMed-2 image segmentation model; a second multileaf grating motion trajectory iteration unit, used to generate a boundary surface based on all three-dimensional coordinates using the Delaunay triangulation algorithm; a third multileaf grating motion trajectory iteration unit, used to set the dose attenuation coefficient according to a preset interval distance, with the center of the boundary surface as the reference point; a fourth multileaf grating motion trajectory iteration unit, used to determine the initial opening and closing state of the multileaf grating and the initial angle of the collimator of the radiotherapy instrument based on the projection contour of the boundary surface; a fifth multileaf grating motion trajectory iteration unit, used to iteratively calculate the optimal opening and closing combination of the multileaf grating at each angle of the collimator using a global optimization algorithm with a preset angle step size, with all optimal opening and closing combinations based on the projection contour that can just completely cover the boundary surface; and a sixth multileaf grating motion trajectory iteration unit, used to convert the multileaf grating position sequence of all optimal opening and closing combinations into multileaf grating motion trajectories along the motion time sequence of the collimator.

[0500] It should be noted that this embodiment is a functional module embodiment based on the above method embodiment. For the preferred, extended, limited, exemplified and principle explanation parts of this embodiment, please refer to the above embodiment. This embodiment will not repeat them.

[0501] This embodiment acquires spectral CT and functional magnetic resonance imaging (fMRI) data of individual patients and generates a fused image matrix dataset using a bilinear registration algorithm. Radiomics features of the fused image matrix dataset are extracted using a ResNet convolutional network. Metabolicly active regions of the radiomics features are identified using Otsu adaptive threshold segmentation to obtain an immune activity feature vector set. An XGBoost classification model is trained using a pre-defined dataset, and the immune activity feature vector set is input into the trained XGBoost model to obtain an immune heatmap. The patient's radiotherapy localization CT scan is acquired and aligned with the immune heatmap coordinate system. The aligned immune heatmap is then segmented at the voxel level using a SAMed-2 image segmentation model, while marking highly responsive voxels. The three-dimensional coordinates of all highly responsive voxels are acquired and integrated into a three-dimensional coordinate set of the immune target area. The motion trajectory of the pre-defined radiotherapy instrument is iterated using Monte Carlo simulation, ensuring that the radiation from the pre-defined radiotherapy instrument covers all three-dimensional coordinates in a decreasing gradient manner. The motion control command sequence after the multi-leaf grating motion trajectory iteration is completed is acquired and input into the pre-defined radiotherapy instrument. This embodiment achieves three-dimensional visualization of metabolic activity and immune response through multimodal fusion of spectral CT / functional magnetic resonance imaging and ResNet radiomics feature extraction. Compared with the identification accuracy of anatomical target volume (GTV), it has a significant improvement. In addition, the immune thermogram generated by the XGBoost model and the SAMed-2 segmentation markers can accurately locate the immune high-response voxels. The gradient dose distribution simulated by Monte Carlo simulation makes the dose intensity of the immune activation area several times that of the conventional area while protecting the inactive area. The multi-leaf grating motion trajectory optimization algorithm of this embodiment can reduce the energy consumption of the lead gate drive of the existing linear accelerator while being compatible with mainstream radiotherapy equipment. Moreover, the XGBoost model training can dynamically adjust the dose hotspot according to the patient's specific immune characteristics, so that the radiation can accurately and sufficiently reach the area in the tumor region that is most likely to activate the immune system.

[0502] Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Figure 3 As shown, the electronic device 8 includes a processor 81 and a memory 82 coupled to the processor 81.

[0503] The memory 82 stores program instructions for implementing the spatial segmentation radiotherapy guidance method based on target immune region activation of any of the above embodiments.

[0504] The processor 81 is used to execute program instructions stored in the memory 82 for spatial segmentation radiotherapy guidance based on target immune region activation.

[0505] The processor 81 can also be referred to as a CPU (Central Processing Unit). The processor 81 may be an integrated circuit chip with signal processing capabilities. The processor 81 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0506] Furthermore, Figure 4 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. See also: Figure 4 In this embodiment of the application, the storage medium 9 stores program instructions 91 capable of implementing all the above methods. These program instructions 91 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0507] In the several embodiments provided in this application, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0508] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A spatial segmentation radiotherapy guidance method based on target area immune region activation, wherein the radiotherapy guidance method is applied to patients undergoing radiotherapy treatment using a pre-set radiotherapy instrument, characterized in that, The radiotherapy guidance method includes: Step S1: Obtain the spectral CT and functional magnetic resonance imaging data of the patient individual, and generate a fused image matrix dataset using a bilinear registration algorithm; Step S2: Extract the radiomics features of the fused image matrix dataset using a ResNet convolutional network; Step S3: Identify the metabolically active regions of the radiomics features using Otsu adaptive threshold segmentation to obtain an immune activity feature vector set; Step S4: Train the XGBoost classification model using a preset dataset based on the XGBoost classification model, and input the immune activity feature vector set into the trained XGBoost classification model to obtain an immune heatmap; Step S5: Obtain the radiotherapy localization CT of the patient and align it with the coordinate system of the immune thermogram. Then, use the SAMed-2 image segmentation model to perform voxel-level segmentation on the aligned immune thermogram and mark the voxels with high immune response. Step S6: Obtain the three-dimensional coordinates of all immune high-response voxels and integrate them into a three-dimensional coordinate set of the immune target area. Iterate the motion trajectory of the multi-leaf grating of the preset radiotherapy instrument through Monte Carlo simulation, and cover all three-dimensional coordinates with the rays of the preset radiotherapy instrument in the form of a preset ray intensity gradient decreasing. Step S7: Obtain the motion control command sequence after the multi-leaf grating motion trajectory iteration is completed and input it into the preset radiotherapy instrument; Step S2, extracting radiomics features from the fused image matrix dataset using a ResNet convolutional network, including: Step S21: Normalize the fused image matrix dataset; Step S22: Divide the normalized fused image matrix dataset into cubes of a preset size; Step S23: Define a ResNet convolutional network and define the number of channels in the input layer of the ResNet convolutional network to match the preset size; Step S24: Input the fused image matrix dataset into the input layer of the ResNet convolutional network, and obtain a multidimensional deep feature vector set through forward propagation of the ResNet convolutional network; Step S25: Filter the top feature vectors in the multidimensional deep feature vector set using the mRMR algorithm; Step S26: All the top feature vectors are upsampled through a deconvolution layer to reconstruct a three-dimensional feature map, and the three-dimensional feature map is defined as the image omics feature; Step S3: Identify the metabolically active regions of the radiomics features using Otsu adaptive threshold segmentation to obtain an immune activity feature vector set, including: Step S31: Normalize the grayscale of the three-dimensional feature map and calculate the histogram distribution of the three-dimensional feature map after grayscale normalization. Step S32: Calculate the inter-class variance of the histogram distribution using the Otsu adaptive threshold; Step S33: Select the optimal threshold among all candidate thresholds that maximizes the inter-class variance, and segment the three-dimensional feature map according to the optimal threshold to form high-metabolic regions and low-metabolic regions. Step S34: Define the high-metabolic region as the metabolically active region; Step S35: Perform morphological closing operation on the metabolically active region based on the preset structural voxel to obtain several connected regions. Step S36: Calculate the gray-level co-occurrence matrix features for each connected region; Step S37: Convert all gray-level co-occurrence matrix features into multi-dimensional feature vectors using a preset Python script; Step S38: Integrate all multidimensional feature vectors and define them as the immune activity feature vector set; Step S5: Obtain the radiotherapy localization CT scan of the patient and align it with the immunothermal map coordinate system. Then, use the SAMed-2 image segmentation model to perform voxel-level segmentation on the aligned immunothermal map, and mark the voxels with high immune response, including: Step S51: Obtain the centroid of the coordinate system of the radiotherapy positioning CT and the centroid of the coordinate system of the immunothermograph; Step S52: Calculate the rotation and translation matrix between the radiotherapy positioning CT and the immunothermograph based on the centroids of the two coordinate systems; Step S53: Based on the rotation and translation matrix, the radiotherapy positioning CT and the immunothermograph are registered using bicubic interpolation to achieve coordinate system alignment; Step S54: The aligned immune heatmap is segmented at the voxel level using the SAMed-2 image segmentation model; Step S55: Mark all voxels in the high-metabolic region as highly immune-responsive voxels.

2. The radiotherapy guidance method according to claim 1, characterized in that, Step S7: Obtain the motion control command sequence after the multi-leaf grating motion trajectory iteration is completed and input it into the preset radiotherapy instrument. Then, the process includes: Step S10: After the radiotherapy treatment is completed by the preset radiotherapy instrument, obtain the SUVmax value matrix of the patient based on FDG-PET / CT scan; Step S20: Calculate the target area FDG uptake change rate based on the baseline SUV and the SUVmax value matrix; Step S30: Determine whether the change rate of FDG uptake in the target area exceeds a preset change rate threshold. If not, proceed to step S40. Step S40: Determine that the motion control command sequence is invalid radiotherapy guidance; In step S50, in response to the invalid radiotherapy guidance, steps S4 to S6 are repeated using an additional preset sample set as the execution subject to continue iteratively updating the XGBoost classification model; Step S60: Obtain the motion control command sequence after further iteration and input it into the preset radiotherapy instrument; Step S70: Repeat steps S10 to S60 until the change rate of FDG uptake in the target area exceeds the preset change rate threshold.

3. The radiotherapy guidance method according to claim 1, characterized in that, Step S4: Train the XGBoost classification model using a preset dataset, and input the immune activity feature vector set into the trained XGBoost classification model to obtain an immune heatmap, including: Step S41: Define the model parameters of the XGBoost classification model based on the preset model strategy; Step S42: Standardize and reduce the dimensions of the preset dataset, and input the sample set in the preset dataset into the XGBoost classification model in batches for training; Step S43: In each batch of training, the XGBoost classification model is validated using the validation set in the preset dataset, and training is terminated when the AUC value of the validation set no longer increases, thus obtaining the XGBoost classification model after training. Step S44: Input the immune activity feature vector set into the trained XGBoost classification model and obtain the positive probability prediction value for each voxel. Step S45: Calculate the numerical range of all positive probability prediction values ​​and assign the numerical range to at least three consecutive and different heat map color levels. Step S46: Map all predicted positive probability values ​​and corresponding heatmap color levels to the coordinate system three-dimensional matrix of the fused image matrix dataset to obtain the immune heatmap.

4. The radiotherapy guidance method according to claim 1, characterized in that, Step S6: Obtain the three-dimensional coordinates of all highly responsive immune voxels and integrate them into a three-dimensional coordinate set of the immune target area. Iterate the motion trajectory of the multi-leaf grating of the preset radiotherapy instrument using Monte Carlo simulation, ensuring that the radiation from the preset radiotherapy instrument covers all three-dimensional coordinates in a preset radiation intensity gradient decreasing manner, including: Step S61: Extract the three-dimensional coordinates of all highly immune voxels based on the immune heatmap segmented by the SAMed-2 image segmentation model. Step S62: Generate the boundary surface based on all three-dimensional coordinates using the Delaunay triangulation algorithm; Step S63: Using the center of the boundary surface as a reference point, set the dose attenuation coefficient according to a preset interval distance; Step S64: Determine the initial opening and closing state of the multi-leaf grating and the initial angle of the collimator of the preset radiotherapy instrument based on the projection profile of the boundary surface. Step S65: Based on the initial angle of the collimator, the optimal opening and closing combination of the multi-leaf grating at each angle of the collimator is calculated iteratively using a global optimization algorithm with a preset angle step size. All optimal opening and closing combinations are based on the projection contour that can just completely cover the boundary surface. Step S66: Using the motion time sequence of the collimator as the chronological order, convert the position sequence of all optimal opening and closing combinations of the multileaf grating into the motion trajectory of the multileaf grating along the motion time sequence.

5. A spatial segmentation radiotherapy guidance system based on target immune region activation, wherein the radiotherapy guidance system is applied to the radiotherapy guidance method as described in any one of claims 1 to 4, characterized in that, The radiotherapy guidance system includes: The fused image matrix dataset acquisition module is used to acquire the spectral CT and functional magnetic resonance data of the patient individual and generate a fused image matrix dataset through a bilinear registration algorithm. The radiomics feature extraction module is used to extract radiomics features from the fused image matrix dataset through a ResNet convolutional network. The immune activity feature vector set acquisition module is used to identify the metabolically active regions of the radiomics features through Otsu adaptive threshold segmentation to obtain the immune activity feature vector set. The immune heatmap acquisition module is used to train the XGBoost classification model based on the XGBoost classification model using a preset dataset, and input the immune activity feature vector set into the trained XGBoost classification model to obtain the immune heatmap. The immune high-response voxel labeling module is used to acquire the radiotherapy positioning CT of the patient and align it with the coordinate system of the immune thermogram. The aligned immune thermogram is then segmented at the voxel level using the SAMed-2 image segmentation model, and immune high-response voxels are labeled simultaneously. The multileaf grating motion trajectory iteration module is used to acquire the three-dimensional coordinates of all immune high-response voxels and integrate them into a three-dimensional coordinate set of the immune target area. Through Monte Carlo simulation, the multileaf grating motion trajectory of the preset radiotherapy instrument is iterated, and the rays of the preset radiotherapy instrument cover all three-dimensional coordinates in the form of a preset ray intensity gradient decreasing. The motion control command sequence output module is used to acquire the motion control command sequence after the multileaf grating motion trajectory iteration is completed and input it into the preset radiotherapy instrument.

6. An electronic device, characterized in that, The method includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the radiotherapy guidance method as described in any one of claims 1 to 4.

7. A storage medium, characterized in that, The storage medium stores program instructions, which, when executed by a processor, enable the radiotherapy guidance method as described in any one of claims 1 to 4.

Images