Radiation therapy position uncertainty risk stratification and adaptive decision making methods and systems

Abstract

The application provides a radiotherapy position uncertainty risk stratification and adaptive decision-making method and system, which comprises the following steps: adopting a deep learning model to perform segmentation processing on a target CT image to obtain a body composition label and an anatomical marker; performing bed removal processing on three-dimensional body surface data, and combining the anatomical marker to determine a region of interest of a patient body surface; extracting a first type of feature from the region of interest of the patient body surface, extracting a second type of feature from the target CT image, and performing feature processing on the first type of feature, the second type of feature and a clinical medical feature to obtain a key feature subset; inputting the key feature subset of a patient to be evaluated into a trained risk stratification model to perform risk stratification, and outputting a risk level in each anatomical direction; and generating an adaptive image verification strategy and an adaptive planning target volume boundary adjustment strategy based on the risk level. Precise prediction of the radiotherapy position uncertainty risk is realized, and a reliable technical tool is provided for clinical implementation.

Description

Technical Field

[0001] This invention belongs to the technical field of radiotherapy for cancer patients, and in particular relates to a method and system for risk stratification and adaptive decision-making regarding radiotherapy positioning uncertainty. Background Technology

[0002] Radiotherapy is one of the main methods of cancer treatment. The accuracy of radiotherapy is crucial to its efficacy and safety. In clinical practice, radiotherapy positioning uncertainty is one of the core factors affecting radiotherapy accuracy. Existing methods for controlling radiotherapy positioning uncertainty mainly rely on CBCT (Cone-Beam Computed Tomography) and SGRT (Surface Guided Radiation Therapy) for successive online validation. However, these methods have the following main shortcomings:

[0003] First, for patients with low risk of radiotherapy positioning uncertainty, unnecessary CBCT scans can increase additional radiation dose and reduce overall radiotherapy efficiency. Second, existing CBCT and SGRT methods cannot quantitatively assess the risk of radiotherapy positioning uncertainty before treatment, i.e., during the radiotherapy planning stage, making it difficult to achieve personalized risk prediction and proactive intervention during the radiotherapy planning stage. Furthermore, the SGRT method can only reflect changes in the external morphology of the patient's body surface and cannot accurately characterize the actual displacement of internal anatomical structures.

[0004] To address the aforementioned shortcomings, existing research has attempted to analyze factors related to radiotherapy positioning uncertainty using a single data source, such as abdominal circumference or body mass index. However, this method has several limitations: first, it is difficult to accurately reflect individual patient differences, resulting in limited predictive accuracy; second, it lacks a multi-dimensional feature fusion mechanism; third, it lacks a clear risk stratification method and a technical pathway that can directly guide clinical decision-making; and fourth, most methods still rely on in-treatment monitoring or post-treatment analysis, lacking pre-treatment predictive capabilities based on individual patient characteristics.

[0005] Therefore, it is necessary to provide a method that can assess the uncertainty risk of radiotherapy positioning based on multimodal information during the radiotherapy planning stage, generate risk stratification results, and produce clinical decision-making basis to guide the implementation of subsequent radiotherapy. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for risk stratification and adaptive decision-making for radiotherapy positioning uncertainty, which solves the problem that existing technologies rely on in-process verification and post-process analysis, but lack risk stratification and clinical decision guidance.

[0007] In a first aspect, the present invention provides a method for risk stratification and adaptive decision-making regarding radiotherapy positioning uncertainty, the method comprising:

[0008] Acquire multimodal data of the patient, including target CT images and three-dimensional body surface data;

[0009] A deep learning model is used to segment the target CT image to obtain body component labels and anatomical markers;

[0010] The three-dimensional body surface data is processed to remove bed material, and combined with the anatomical markers, the region of interest on the patient's body surface is determined.

[0011] The first type of features are extracted from the region of interest on the patient's body surface, and the second type of features are extracted from the target CT image. The first type of features, the second type of features, and the clinical medical features are then processed to obtain a subset of key features.

[0012] A subset of key features of the patient to be assessed is input into a trained risk stratification model to perform risk stratification and output the risk level in each anatomical direction.

[0013] Based on the risk level, an adaptive image verification strategy and an adaptive planned target area boundary adjustment strategy are generated.

[0014] In one implementation, the step of debedring the three-dimensional body surface data and, in conjunction with the anatomical markers, determining the region of interest on the patient's body surface includes:

[0015] Spatial morphological processing is performed on the three-dimensional body surface data to obtain primary body surface data;

[0016] A back reference plane is established based on the anatomical features of the primary body surface data, and interfering data below the back reference plane is removed to obtain the corrected body surface data.

[0017] The corrected body surface data is extracted along the long axis of the human body to obtain the purified body surface data;

[0018] Based on the purified body surface data and the anatomical markers, the region of interest on the patient's body surface is determined.

[0019] In one implementation, before performing feature processing on the first type of features, the second type of features, and the clinical medical features, the following steps are included:

[0020] Historical radiotherapy positioning error data for each patient was obtained, including positioning error measurements in each anatomical direction.

[0021] Based on the historical radiotherapy positioning error data, the mean and standard deviation of the absolute values ​​of positioning errors in each anatomical direction for each historical patient were calculated.

[0022] Based on the mean, the standard deviation, and the preset risk assessment threshold, risk labels for each anatomical direction of each historical patient are generated.

[0023] In one implementation, generating risk labels for each anatomical direction for each historical patient based on the mean, the standard deviation, and a preset risk assessment threshold includes: each anatomical direction includes left-right, head-to-foot, and front-to-back directions, wherein...

[0024] For the left-right and front-back directions, a first judgment threshold is calculated based on the mean and standard deviation. If the first judgment threshold is greater than the first preset risk judgment threshold, the corresponding direction is judged as high risk; otherwise, it is judged as low risk.

[0025] For the head-to-toe direction, a second judgment threshold is calculated based on the mean and standard deviation. If the second judgment threshold is greater than the second preset risk judgment threshold, the head-to-toe direction is judged as high risk; otherwise, it is judged as low risk.

[0026] In one implementation, the training method of the risk stratification model includes:

[0027] Using the risk labels as supervisory signals, the key feature subsets corresponding to each historical patient are input into a machine learning method for training to obtain the risk stratification model.

[0028] In one implementation, generating an adaptive image verification strategy and an adaptive target area boundary adjustment strategy based on the risk level includes:

[0029] The adaptive image verification strategy includes: counting the number of high-risk directions in each anatomical direction based on the risk level of the patient to be evaluated in each anatomical direction, and determining the initial CBCT verification frequency based on the number of high-risk directions;

[0030] The initial CBCT verification frequency is dynamically adjusted based on the radiotherapy positioning error data obtained from online CBCT. This dynamic adjustment includes: acquiring radiotherapy positioning error data for each treatment fraction and calculating the absolute value of the radiotherapy positioning error data for each anatomical direction; monitoring the number of times the radiotherapy positioning error data for each anatomical direction exceeds a corresponding preset error threshold within a preset observation window; when the number of times the radiotherapy positioning error data for at least one anatomical direction exceeds the corresponding preset error threshold reaches a preset threshold, it is determined that the patient to be evaluated has a persistent positioning risk in the corresponding anatomical direction, and the initial CBCT verification frequency is increased to a second CBCT verification frequency.

[0031] When the online radiotherapy positioning error data of each anatomical direction of the patient to be evaluated is lower than the corresponding recovery threshold in a continuous preset number of radiotherapy fractions, the risk level is updated according to the radiotherapy positioning error data of the existing CBCT fractions in the radiotherapy cycle, and the second CBCT verification frequency is restored to the CBCT verification frequency corresponding to the updated risk level.

[0032] The adaptive planning target boundary adjustment strategy includes: classifying each historical patient into a low-risk group and a high-risk group based on the risk labels of each historical patient in each anatomical direction;

[0033] Acquire first historical radiotherapy positioning error data for historical patients in the low-risk group in each anatomical direction, and determine first systematic error data and first random error data for historical patients in the low-risk group in each anatomical direction based on the first historical radiotherapy positioning error data; acquire second historical radiotherapy positioning error data for historical patients in the high-risk group in each anatomical direction, and determine second systematic error data and second random error data for historical patients in the high-risk group in each anatomical direction based on the second historical radiotherapy positioning error data;

[0034] Based on the first systematic error data and the first random error data, the van Herk formula is used to calculate the first planned target area expansion boundary of the low-risk group in each anatomical direction; based on the second systematic error data and the second random error data, the van Herk formula is used to calculate the second planned target area expansion boundary of the high-risk group in each anatomical direction.

[0035] Based on the risk level of each patient under evaluation in each anatomical direction, the patients under evaluation are divided into a low-risk group and a high-risk group. For the patients under evaluation in the low-risk group, the clinical target area of ​​the patients under evaluation is expanded outward in the corresponding anatomical direction based on the outer boundary of the first planned target area to obtain the first planned target area. For the patients under evaluation in the high-risk group, the clinical target area of ​​the patients under evaluation is expanded outward in the corresponding anatomical direction based on the outer boundary of the second planned target area to obtain the second planned target area. Both the first planned target area and the second planned target area are not lower than the minimum clinical safety threshold.

[0036] In one embodiment, before performing feature processing on the first type of features, the second type of features, and the clinical medical features, the method further includes:

[0037] The continuous features in the first type of features, the second type of features, and clinical medical features are truncated at 1% to 99% quantiles, and feature values ​​that exceed the 1% to 99% quantile range are replaced with corresponding boundary feature values.

[0038] The first type of features, the second type of features, and the multi-category nominal features in clinical medical features are converted into 0 or 1 integer vectors using one-hot encoding.

[0039] The first type of features, the second type of features, and clinical medical features are standardized.

[0040] The first type of features includes at least multidimensional body surface morphological features, and the second type of features includes at least radiomics features, spatial anatomical relationship features, and body composition quantification features.

[0041] In one implementation, feature processing is performed on the first type of features, the second type of features, and clinical medical features to obtain a subset of key features, including:

[0042] Features whose variance is less than a first preset threshold among the first type of features, the second type of features, and clinical medical features are removed to obtain the first combined features;

[0043] The first combined features are filtered using the Pearson correlation coefficient to obtain the second combined features;

[0044] The second combined features are sorted in descending order of importance score, and the embedded feature selection method is used to filter the second combined features to obtain a subset of key features.

[0045] In one embodiment, the anatomical markers include a first anatomical marker, a second anatomical marker, and a third anatomical marker, and the method for calculating the body composition quantification characteristics is as follows:

[0046] Using the cardiac level of the first anatomical marker material as the upper boundary and the cardiac level of the second anatomical marker material as the lower boundary, the target CT image is cropped to determine the range of three-dimensional volume composition analysis.

[0047] The cardiac level, the third anatomical marker, is used as the scope for two-dimensional body composition analysis.

[0048] The volume composition quantification features are obtained based on the two-dimensional volume composition analysis range and the three-dimensional volume composition analysis range.

[0049] In one embodiment, the multidimensional surface morphological features include a set of micro-statistical features reflecting the statistical distribution of point cloud shape index and curvature, a set of global geometric features reflecting the overall geometric morphology, a set of local structural features reflecting the distribution of local curvature and normal, and a set of topological features reflecting the topological integrity of the mesh.

[0050] The image omics features include geometric shape descriptors, voxel gray value statistics, and texture descriptors;

[0051] The spatial anatomical relationship features include: the minimum three-dimensional Euclidean distance from the centroid of the target radiotherapy area to the body surface; the minimum three-dimensional distance from the centroid of the target radiotherapy area to the bone; the minimum three-dimensional Euclidean distance from the isocenter to the body surface; the minimum three-dimensional distance from the isocenter to the bone; the three-dimensional Euclidean distance from the isocenter to the centroid of the target radiotherapy area and the components of the three-dimensional Euclidean distance in each direction; and, on the cross-section passing through the isocenter, the ratio of the anteroposterior diameter to the left-right diameter and the perimeter of the body surface contour at the isocenter level.

[0052] In one implementation, before segmenting the target CT image using a deep learning model, the process includes:

[0053] The spatial mapping relationship between the three-dimensional body surface data and the target CT image is obtained, and the three-dimensional body surface data is spatially aligned based on the spatial mapping relationship to obtain aligned three-dimensional body surface data.

[0054] In one embodiment, the body composition label includes at least subcutaneous fat, visceral fat, muscle tissue, intramuscular fat, and bone tissue.

[0055] Secondly, this application provides a risk stratification and adaptive decision-making system for radiotherapy positioning uncertainty, the system comprising:

[0056] The data acquisition module is configured to acquire multimodal data of the patient, including target CT images and three-dimensional body surface data;

[0057] The data processing module is configured to use a deep learning model to segment the target CT image to obtain body component labels and anatomical markers;

[0058] The region of interest generation module is configured to perform bed removal processing on the three-dimensional body surface data and, in conjunction with the anatomical markers, determine the region of interest on the patient's body surface.

[0059] The feature processing module is configured to extract a first type of feature from the region of interest on the patient's body surface, extract a second type of feature from the target CT image, and perform feature processing on the first type of feature, the second type of feature, and clinical medical features to obtain a subset of key features.

[0060] The risk prediction module is configured to input a subset of key features of the patient to be assessed into a trained risk stratification model to perform risk stratification and output the risk level in each anatomical direction.

[0061] The adaptive decision-making module is configured to generate an adaptive image verification strategy and an adaptive target area boundary adjustment strategy based on the risk level.

[0062] As described above, the radiotherapy positioning uncertainty risk stratification and adaptive decision-making method and system of the present invention have the following beneficial effects:

[0063] First, by employing a deep learning model to automatically segment CT images and combining it with bed removal processing, regions of interest reflecting the true anatomical structure can be accurately identified, effectively eliminating non-physiological interferences such as the treatment bed and providing a high-quality data foundation for subsequent analysis. Second, by integrating surface morphological features, radiomics features, and spatial anatomical relationship features to construct a multi-dimensional feature system, it can comprehensively capture the biomechanical characteristics of positional stability (such as subcutaneous fat distribution) and spatial constraints (such as the minimum three-dimensional distance from the target radiotherapy area contour to bone tissue), improving the clinical relevance of key feature subsets. Finally, based on the multi-directional risk grading output of the machine learning model, potential positioning risks in various anatomical directions can be specifically identified. Compared with traditional single-dimensional assessment methods, this provides a reliable technical tool for positional management in clinical practice to achieve individualized and precise radiotherapy, and provides a quantitative basis for personalized positioning fixation and image-guided frequency adjustment. Attached Figure Description

[0064] Figure 1 A flowchart of a risk stratification and adaptive decision-making method for radiotherapy positioning uncertainty is provided in this embodiment of the invention;

[0065] Figure 2 A schematic diagram of the two-dimensional volume composition analysis range including volume composition labels provided in an embodiment of the present invention;

[0066] Figure 3 A three-dimensional view of the target CT image provided in an embodiment of the present invention;

[0067] Figure 4 This is a schematic diagram of the sagittal plane of the target CT image provided in an embodiment of the present invention;

[0068] Figure 5 A flowchart of a method for determining the region of interest on a patient's body surface by performing bed removal processing on the three-dimensional body surface data and combining it with the anatomical markers, provided in an embodiment of the present invention;

[0069] Figure 6 A flowchart of a method for generating regions of interest on the body surface provided in an embodiment of the present invention;

[0070] Figure 7 A flowchart of a method for generating risk labels for each anatomical direction of historical patients for risk stratification model training, provided in an embodiment of the present invention;

[0071] Figure 8 A schematic diagram illustrating the spatial relationship between isocenters, target radiotherapy zone, body surface, and bones based on target CT images, provided in an embodiment of the present invention.

[0072] Figure 9 This is a cross-sectional schematic diagram of the isocenter points provided in an embodiment of the present invention;

[0073] Figure 10 A flowchart of a method for performing feature processing on first-type features, second-type features, and clinical medical features to obtain a subset of key features, provided in an embodiment of the present invention;

[0074] Figure 11 A flowchart of a method for generating an adaptive image verification strategy based on risk level, provided in an embodiment of the present invention;

[0075] Figure 12 A flowchart illustrating a method for generating an adaptive target area boundary adjustment strategy based on risk level, as provided in an embodiment of the present invention.

[0076] Figure 13 A flowchart of a radiotherapy position uncertainty risk stratification and adaptive decision-making method is provided as another exemplary embodiment of the present invention;

[0077] Figure 14 A structural diagram of a radiotherapy position uncertainty risk stratification and adaptive decision-making system provided in an embodiment of the present invention;

[0078] Figure 15 This is a block diagram of an electronic device.

[0079] Explanation of reference numerals in the accompanying drawings: 1. Data acquisition module; 2. Data processing module; 3. Region of interest generation module; 4. Feature processing module; 5. Risk prediction module; 6. Adaptive decision-making module; 20. Electronic device; 21. Processor; 22. Memory; 23. Output interface; 24. Communication interface; 25. Antenna. Detailed Implementation

[0080] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0081] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0082] To address the aforementioned technical issues, this invention proposes a method and system for risk stratification and adaptive decision-making regarding radiotherapy positioning uncertainty. This method enables automated, high-precision, quantifiable, and directional prediction of radiotherapy positioning uncertainty risk, providing a reliable technical tool for positioning management in clinical practice to achieve individualized and precise radiotherapy.

[0083] The technical solutions of the present invention will now be described in detail with reference to the accompanying drawings.

[0084] like Figure 1 As shown in the figure, this embodiment provides a method for risk stratification and adaptive decision-making regarding radiotherapy positioning uncertainty. The method includes:

[0085] Step 100: Acquire multimodal data of the patient, including target CT images and three-dimensional body surface data.

[0086] The patients include those to be evaluated and those with a history of radiation therapy. The history of radiation therapy is the sample population used to train the risk stratification model. The patients to be evaluated are those for whom the trained risk stratification model will be used to assess the risk of radiation therapy positioning errors. These include patients who have not yet received their first radiation therapy, patients who have received radiation therapy but have not yet started their current treatment, and patients whose current radiation therapy treatment has not yet been fully completed.

[0087] In some embodiments, the multimodal data includes, but is not limited to, the patient's clinical medical data, target CT images, three-dimensional body surface data, and historical radiotherapy positioning errors. It should be noted that only historical patients require the collection of this historical radiotherapy positioning error data; patients to be evaluated do not require this historical radiotherapy positioning error data.

[0088] Specifically, the clinical medical data includes, but is not limited to, measurements such as height, weight, sex, age, body mass index, body fat percentage, Eastern Cooperative Oncology Group (ECOG) score, radiation therapy site and whether an abdominal depressor was used, and numerical rating scale (NRS). Historical radiotherapy positioning error data can be obtained via CBCT, and the target CT images are extracted using X-ray scanning and computer reconstruction techniques on a CT scanner.

[0089] In this embodiment, the three-dimensional body surface data is generated from the target CT image to create a reference body surface. Specifically, a threshold segmentation algorithm or edge detection algorithm is used to extract the patient's body contour from the target CT image to generate a reference body surface model. Since the reference body surface model is directly derived from the target CT image, it has the same spatial coordinates as the target CT image, meaning they are already in the same coordinate system. At this point, an identity mapping relationship is established, and no additional rotation or translation transformation is required. The reference body surface model is directly used as the three-dimensional body surface data for subsequent feature extraction.

[0090] In some embodiments, the three-dimensional body surface data includes at least three-dimensional body surface point cloud data or three-dimensional body surface mesh data.

[0091] Step 200: Use a deep learning model to segment the target CT image to obtain body component labels and anatomical markers.

[0092] Specifically, a deep learning model is used to automatically segment the target CT image into multiple tissues to obtain volume component labels, such as... Figure 2 As shown, the body composition label includes, but is not limited to, subcutaneous fat, visceral fat, muscle tissue, intramuscular fat, and bone tissue.

[0093] In this embodiment, nnU-Net is used to automatically segment the target CT image into multiple tissues to obtain volume component labels. nnU-Net (no-new-Net) is a self-configurable medical image segmentation framework that can automatically determine the network structure (based on the U-Net backbone), preprocessing strategy, and training hyperparameters according to the characteristics of the dataset. It has achieved leading levels in multiple medical image segmentation benchmark tests and is highly compatible with the multi-tissue fine segmentation requirements of this invention.

[0094] Specifically, a deep learning model is used to automatically segment the anatomical markers of the target CT image to obtain the three-dimensional volume composition analysis range and the two-dimensional volume composition analysis range. The anatomical markers include a first anatomical marker, a second anatomical marker, and a third anatomical marker. Specifically, such as... Figure 3 and Figure 4 As shown, the target CT image is cropped with the first anatomical landmark at the cardiac level as the upper boundary and the second anatomical landmark at the cardiac level as the lower boundary to determine the three-dimensional volume composition analysis range; the third anatomical landmark at the cardiac level is used as the two-dimensional volume composition analysis range. In this embodiment, the first anatomical landmark is the T12 vertebra, the second anatomical landmark is the L4 vertebra, and the third anatomical landmark is the L3 vertebra.

[0095] In this embodiment, TotalSegmentor is used to automatically segment the target CT image using anatomical markers to obtain the three-dimensional volume composition analysis range and the two-dimensional volume composition analysis range. TotalSegmentor is built on the nnU-Net framework and has been pre-trained on a large-scale multi-center CT image dataset. It supports fully automatic segmentation of more than 100 anatomical structures and can directly output the anatomical markers required by this invention. It has out-of-the-box segmentation capabilities, which significantly reduces the threshold for clinical deployment.

[0096] It should be noted that deep learning models are not limited to the specific implementations described above; any model capable of performing the corresponding segmentation task is within the scope of protection of this invention.

[0097] In some embodiments, the target CT image includes, but is not limited to, body contours, the target radiotherapy volume, and isocenter coordinates. The target radiotherapy volume includes, but is not limited to, the gross tumor volume (GTV), clinical target volume (CTV), and planning target volume (PTV). In this embodiment, the target radiotherapy volume preferably uses the CTV. This embodiment employs a deep learning model for automatic segmentation of the target CT image, specifically extracting anatomical landmarks, improving positioning accuracy, reducing the standardization error of the analysis range, and cropping the target CT image based on anatomical landmarks, thus improving the topological integrity of the body surface data. Furthermore, the body component quantification features in this embodiment are computed in parallel, greatly reducing the time required to generate body component quantification features. This solves the problem of large radiotherapy positioning errors caused by the instability of the anatomical reference frame in existing technologies.

[0098] Step 300: Perform bed removal processing on the three-dimensional body surface data and combine it with the anatomical markers to determine the region of interest on the patient's body surface.

[0099] Specifically, such as Figure 5 and Figure 6 As shown, the three-dimensional body surface data was debed, and combined with the anatomical landmarks, the region of interest on the patient's body surface was determined, including:

[0100] Step 301: Perform spatial morphological processing on the three-dimensional body surface data to obtain primary body surface data.

[0101] In this embodiment, the spatial morphology processing includes a morphological dilation operation.

[0102] In some embodiments, such as Figure 6As shown, the morphological dilatation operation performs isotropic dilatation in each anatomical direction with a dilatation radius of 6mm to 10mm. In this embodiment, the dilatation radius for isotropic dilatation in each anatomical direction is 8mm. It should be noted that the dilatation radius for isotropic dilatation in each anatomical direction can be appropriately adjusted according to the CT resolution and the patient's body shape to ensure that the patient's three-dimensional body surface point cloud is fully covered without excessively including areas far from the body surface, generating an extended mask. This extended mask is used to filter the initial mesh data after the three-dimensional body surface point cloud data is meshed, to separate the body surface data from the bed board data, and to remove data points located outside the patient's body area to obtain primary body surface data.

[0103] In some embodiments, when the initial grid data after the three-dimensional body surface point cloud data is not available, the segmentation label can be reconstructed based on the target CT image to generate an equivalent segmentation label, and then spatial morphological processing can be performed.

[0104] Step 302: Establish a back reference plane based on the anatomical features of the primary body surface data, and remove interfering data below the back reference plane to obtain the corrected body surface data.

[0105] Specifically, such as Figure 6 As shown, the back reference plane is the plane on the body surface that is in close contact with the bed board. Based on this back reference plane, the back surface is flattened along the direction of the bed board to remove the bed board structure that is in close contact with the back. In this embodiment, the direction of the bed board corresponds to the Z-axis direction of the radiotherapy coordinate system and the Y-axis direction of the CT image coordinate system.

[0106] In some embodiments, the smoothing thickness of the back surface is 30mm to 60mm. In this embodiment, the smoothing thickness is 50mm. It should be noted that this smoothing thickness can be determined based on the actual thickness of the bed board and the contact area of ​​the patient's back surface.

[0107] Step 303: Extract the corrected body surface data along the long axis of the human body to obtain the purified body surface data.

[0108] Specifically, such as Figure 6 As shown, the corrected body surface data is cropped along the long axis of the human body, with a cropping range of 1-3 mm. In this embodiment, the cropping amount is 2 mm for the head side and 2 mm for the foot side. In this embodiment, the long axis of the human body corresponds to the Z-axis direction in the CT image coordinate system.

[0109] This embodiment employs bed removal processing on the three-dimensional body surface data, effectively eliminating non-physiological artifacts and interference data introduced by fixation devices such as treatment bed boards. This reduces the error in body surface data extraction and the time spent on bed removal processing, establishing a precise data foundation for the subsequent extraction accuracy of regions of interest on the patient's body surface.

[0110] It should be noted that the values ​​of the above parameters are determined based on the actual clinical bed size, CT scan characteristics and patient body shape distribution experience. They can also be adaptively adjusted according to the specific equipment model and patient group characteristics. Any combination of parameters that can achieve the equivalent bed removal effect or the generation effect of region of interest is within the protection scope of this invention.

[0111] Step 304: Based on the purified body surface data and the anatomical markers, determine the region of interest on the patient's body surface.

[0112] Specifically, based on the purified body surface data and the three-dimensional body composition analysis range determined by the anatomical markers, the region of interest on the patient's body surface is determined.

[0113] like Figure 6 As shown, the purified surface data in the Z-axis direction extracted in step 303 is cropped based on the boundary of the region of interest (i.e., the range of 3D volume composition analysis), retaining all valid face elements whose three vertices are located within the boundary of the region of interest, thus generating the final region of interest.

[0114] This embodiment uses precise anatomical markers and purified body surface data to jointly determine the region of interest on the patient's body surface, ensuring a high degree of anatomical consistency and repeatability of the feature extraction region. It avoids the shift of the region of interest caused by bed board artifacts or segmentation errors, and enables effective comparison of features between different patients and between different treatments of the same patient.

[0115] Step 400: Extract the first type of features from the region of interest on the patient's body surface and the second type of features from the target CT image. Perform feature processing on the first type of features, the second type of features, and clinical medical features to obtain a subset of key features.

[0116] In some embodiments, the clinical medical characteristics include at least height, weight, sex, age, body mass index, body fat percentage, Eastern Cooperative Oncology Group (ECOG) score, radiation therapy site and whether an abdominal pressure plate was used, and numerical rating scale (NRS).

[0117] In some embodiments, the anatomical markers include a first anatomical marker, a second anatomical marker, and a third anatomical marker. The method for calculating the body composition quantification features is as follows: the target CT image is cropped with the cardiac level of the first anatomical marker as the upper boundary and the cardiac level of the second anatomical marker as the lower boundary to determine the three-dimensional body composition analysis range; the cardiac level of the third anatomical marker is used as the two-dimensional body composition analysis range, and the body composition quantification features are obtained based on the two-dimensional and three-dimensional body composition analysis ranges. These body composition quantification features include, but are not limited to, the ratio of visceral fat to subcutaneous fat, the ratio of intramuscular fat to muscle, the skeletal muscle mass index (SMI), and the visceral adiposity index (VAI).

[0118] In some embodiments, such as Figure 7 As shown, before performing feature processing on the first type of feature, the second type of feature, and the clinical medical feature, the following steps are included:

[0119] Step 411: Obtain historical radiotherapy positioning error data for each historical patient. This error data includes positioning error measurements for each anatomical direction.

[0120] Step 412: Based on the historical radiotherapy positioning error data, calculate the mean and standard deviation of the absolute values ​​of the positioning errors in each anatomical direction for each historical patient.

[0121] Step 413: Based on the mean, the standard deviation, and the preset risk assessment threshold, generate risk labels for each anatomical direction of each historical patient.

[0122] It should be noted that steps 411 to 413 are used to generate the risk labels required for the risk stratification model training phase. These are preparatory steps completed offline in advance during the construction of the risk stratification model trained in step 500, and are implemented based on the historical treatment data accumulated by each patient. For the patient to be evaluated, it is not necessary to have their own historical radiotherapy positioning error data. Steps 100 to 500 need to be executed sequentially to obtain the risk level prediction results of the patient to be evaluated in each anatomical direction, thereby achieving a truly personalized pre-treatment risk assessment.

[0123] Specifically, based on the mean, the standard deviation, and a preset risk assessment threshold, risk labels for each anatomical direction of each historical patient are generated. These anatomical directions include left-right, head-to-toe, and front-to-back directions. For the left-right and front-to-back directions, a first assessment threshold is calculated based on the mean and the standard deviation. If the first assessment threshold is greater than the first preset risk assessment threshold, the corresponding direction is assessed as high-risk; otherwise, it is assessed as low-risk. For the head-to-toe direction, a second assessment threshold is calculated based on the mean and the standard deviation. If the second assessment threshold is greater than the second preset risk assessment threshold, the head-to-toe direction is assessed as high-risk; otherwise, it is assessed as low-risk.

[0124] Furthermore, this anatomical orientation includes the patient's left-right direction (X-axis), head-to-foot direction (Y-axis), and anterior-posterior direction (Z-axis). It should be noted that these directions refer to the X-axis, Y-axis, and Z-axis in the radiotherapy coordinate system. Specifically, the X-axis in the radiotherapy coordinate system corresponds to the X-axis in the CT image coordinate system, the Y-axis corresponds to the Z-axis in the CT image coordinate system, and the Z-axis corresponds to the Y-axis in the CT image coordinate system. It should be noted that the radiotherapy coordinate system is defined according to IEC 61217 (equivalent to the Chinese national standard GB / T 18987 "Coordinate Systems, Motion and Scales of Radiotherapy Equipment"), while the CT image coordinate system is defined according to the DICOM standard.

[0125] In some embodiments, the first preset risk assessment threshold and the second preset risk assessment threshold are determined based on the statistical distribution of historical radiotherapy positioning error data.

[0126] For example, for the patient's left-right direction (X-axis) and front-back direction (Z-axis), the first judgment threshold is the mean of the absolute values ​​of the radiotherapy positioning errors in the left-right or front-back directions plus N times the standard deviation, where N is a real number greater than 0. In this embodiment, N is 2, that is, the first judgment threshold is the mean of the absolute values ​​of the radiotherapy positioning errors in the left-right or front-back directions plus 2 times the standard deviation. For the patient's head-to-foot direction (Y-axis), the second judgment threshold is the mean of the absolute values ​​of the radiotherapy positioning errors in the head-to-foot direction plus N times the standard deviation, where N is a real number greater than 0. In this embodiment, N is 2, that is, the second judgment threshold is the mean of the absolute values ​​of the radiotherapy positioning errors in the head-to-foot direction plus 2 times the standard deviation. Meanwhile, the first preset risk judgment threshold is 0.7cm, and the second preset risk judgment threshold is 1.0cm. That is, when the calculated first judgment threshold is greater than 0.7cm, the corresponding direction is judged as high risk, otherwise it is judged as low risk; when the calculated second judgment threshold is greater than 1.0cm, the head-to-toe direction is judged as high risk, otherwise it is judged as low risk.

[0127] In some embodiments, before performing feature processing on the first type of feature, the second type of feature, and the clinical medical feature, the method further includes: truncating the continuous features in the first type of feature, the second type of feature, and the clinical medical feature by 1% to 99th percentile reduction, and replacing feature values ​​exceeding the 1% to 99th percentile range with their corresponding quantile values. Specifically, feature values ​​below the 1% quantile are replaced with the 1% quantile, and feature values ​​above the 99th percentile are replaced with the 99th percentile. For example, assuming the boundary feature value corresponding to the 99th percentile is 1000, when an extreme feature value of 10000 appears, the extreme feature value 10000 exceeding the 99th percentile is directly replaced with the boundary feature value 1000 corresponding to the 99th percentile. The multi-category nominal features in the first type of feature, the second type of feature, and the clinical medical feature are converted into 0 or 1 integer vectors using one-hot encoding to avoid introducing spurious order relationships in numerical encoding; in some embodiments, the multi-category nominal features include, but are not limited to, tumor location. For binary features in this multi-category nominal feature, such as gender, the original values ​​of 0 or 1 are directly retained. The first type feature, the second type feature, and the clinical medical feature are standardized to ensure that they are all of the same magnitude, i.e., the mean of the first type feature, the second type feature, and the clinical medical feature is 0 and the standard deviation is 1, thereby eliminating the influence of different dimensions.

[0128] In this embodiment, the first type of feature includes at least multidimensional body surface morphological features, and the second type of feature includes at least radiomics features, spatial anatomical relationship features, and body composition quantification features. The normalization process is Z-score processing.

[0129] In some embodiments, those skilled in the art can also replace other standardization processing methods as needed, such as min-maximum normalization. This embodiment preprocesses the multidimensional body surface morphology features, the radiomics features, and the spatial anatomical relationship features. By unifying the feature scale, eliminating data noise, and standardizing the data structure, it significantly improves the stability of model training and the accuracy of prediction, while ensuring the comparability of different modal features during fusion, providing a standardized, high signal-to-noise ratio input feature set for subsequent risk stratification.

[0130] Specifically, the multidimensional surface morphological features include a set of micro-statistical features reflecting the statistical distribution of point cloud shape index and curvature, a set of global geometric features reflecting the overall geometric shape, a set of local structural features reflecting the distribution of local curvature and normal, and a set of topological features reflecting the integrity of the mesh topology.

[0131] Furthermore, this set of microscopic statistical features includes a family of shape index statistics and a family of curvature statistics. The shape index statistics family includes, but is not limited to, entropy, mean, standard deviation, skewness, kurtosis, median, interquartile range, 10th percentile, 90th percentile, total energy, root mean square, mean absolute deviation, robust mean absolute deviation, maximum value, uniformity, and range, used to characterize the distribution of surface undulations. The curvature statistics family also includes features from the aforementioned shape index statistics family, characterizing the sharpness of surface undulations. Additionally, the mean and standard deviation of surface roughness are calculated as proxy indicators for respiratory artifacts. This set of global geometric features includes, but is not limited to, principal axis scale, overall isotropy, overall length, body surface area, bed planar projected area, surface elevation ratio, estimated volume, sphericity, left-right symmetry, surface area to volume ratio, normal vector isotropy, normal vector flatness, and normal vector linearity, used to comprehensively characterize the three-dimensional scale, shape, symmetry, and surface-volume relationship of the torso. The local structural feature group, divided into anatomical region proportions based on shape indices, includes ridge proportion, valley proportion, bulge proportion, and pit proportion. Specifically, the ridge proportion can range from [0.375, 0.625], the valley proportion from [-0.625, -0.375], the bulge proportion from [0.875, 1.0], and the pit proportion from [-1.0, -0.875]. The pit morphology features include pit depth and average pit curvature, totaling six features, corresponding to anatomical landmarks such as ribs, grooves, abdominal protrusions, navels, and scars. The topological feature group includes seven features: number of holes, hole area percentage, maximum hole area, inverse of hole perimeter ratio, hole Z-axis centroid, average hole edge curvature, and bulge filling rate. These features are used to assess SGRT data coverage, locate the head-to-foot position of occluded areas, and predict the risk of surface registration failure.

[0132] Specifically, the image omics features include, but are not limited to, geometric shape descriptors, voxel gray value statistics, and texture descriptors.

[0133] In some embodiments, the radiomics features are extracted based on the target CT image. When extracting radiomics features based on the target CT image, for body contours, subcutaneous fat, visceral fat, muscle tissue, intramuscular fat and bone tissue, the corresponding features are calculated only within the scope of their three-dimensional volume composition analysis. For the target radiotherapy target area, the radiomics features are extracted separately, not limited to the above-mentioned three-dimensional volume composition analysis scope.

[0134] Furthermore, the geometric descriptor includes, but is not limited to, the volume, surface area, sphericity, compactness, and ratio of major to minor axes of the region of interest. The voxel grayscale statistics include, but are not limited to, the mean, median, standard deviation, skewness, kurtosis, energy, and entropy of the voxel grayscale values ​​of the region of interest. The texture descriptor includes, but is not limited to, those based on the gray-level co-occurrence matrix, gray-level run-length matrix, and gray-level region size matrix. Specifically, such as... Figure 8 As shown, the spatial anatomical relationship features include: the minimum three-dimensional Euclidean distance from the centroid of the target radiotherapy area to the body surface; the specific calculation formula is as follows: ,in,( , , ) represents the coordinates of the centroid of the target radiotherapy area. , , ( ) represents all points on the body surface. This embodiment quantifies the spatial buffer distance between the body surface and the centroid of the target radiotherapy area by calculating the minimum three-dimensional Euclidean distance from the centroid of the target radiotherapy area to the body surface. This is used to predict the risk of radiotherapy positioning error transmission; that is, the smaller the distance from the centroid of the target radiotherapy area to the body surface, the greater the impact of body surface displacement on the target radiotherapy area. The minimum three-dimensional distance from the centroid of the target radiotherapy area to the bone; the specific calculation formula is as follows: in,( , , () represents all points on the bone. This embodiment quantifies the spatial coupling between the centroid of the target radiotherapy area and the bone by calculating the minimum three-dimensional distance from the centroid of the target radiotherapy area to the bone. That is, the smaller the distance from the centroid of the target radiotherapy area to the bone, the more the bone, as an internal reference, reduces target displacement. The minimum three-dimensional Euclidean distance from the isocenter point to the body surface; the specific calculation formula is as follows: in,( , , The coordinates of the isocenter are shown below. This embodiment monitors the proximity of the isocenter to the body surface by calculating the minimum three-dimensional Euclidean distance from the isocenter to the body surface, preventing a decrease in the tolerance for radiotherapy positioning due to an excessively small distance between the isocenter and the body surface. The minimum three-dimensional distance from the isocenter to the bone is also shown below; the specific calculation formula is as follows: This embodiment establishes a bone-based isocenter position verification benchmark by calculating the minimum three-dimensional distance from the isocenter to the bone. The three-dimensional Euclidean distance from the isocenter to the centroid of the target radiotherapy area is calculated using the following formula: ,in,( , , The coordinates of the centroid of the target radiotherapy region are given, and these coordinates are the geometric mean of all points within the target radiotherapy region. This embodiment directly evaluates the accuracy of radiotherapy positioning by calculating the three-dimensional Euclidean distance from the isocenter point to the centroid of the target radiotherapy region. For example... Figure 9 As shown, on a cross-section passing through the isocenter point, calculate the ratio of the anteroposterior diameter to the lateral diameter, and the perimeter of the surface contour at the isocenter point level. The specific calculation formula is as follows: Circumference= ,in( , () represents the surface contour point of the body at the center level on the cross-section.

[0135] It should be noted that all the above calculations are based on the physical dimensions in the CT image coordinate system, and the coordinate consistency of the multimodal data is ensured through a spatial transformation matrix.

[0136] Specifically, such as Figure 10 As shown, feature processing is performed on the first type of feature, the second type of feature, and the clinical medical feature to obtain a subset of key features, including:

[0137] Step 421: Remove the first type of feature, the second type of feature, and the clinical medical feature whose variance is less than the first preset threshold to obtain the first combined feature.

[0138] In this embodiment, the first type of feature includes at least multidimensional body surface morphological features, and the second type of feature includes at least radiomics features, spatial anatomical relationship features, and body composition quantification features. Specifically, the variances of the multidimensional body surface morphological features, radiomics features, spatial anatomical relationship features, clinical medical features, and body composition quantification features on the training set are calculated, and features with variances lower than a first preset threshold are removed. It should be noted that the acquisition of the first combination of features does not depend on risk labels and is an unsupervised filtering process.

[0139] In some embodiments, the first preset threshold can be set by those skilled in the art based on clinical trial validation. In this embodiment, the first preset threshold is 0.01.

[0140] Step 422: Filter the first combination of features using the Pearson correlation coefficient to obtain the second combination of features.

[0141] Specifically, the Pearson correlation coefficient between any two features in the first combination of features is calculated. Filter out those that meet the requirements Feature pairs with a value ≥0.80, and retain For feature pairs with a correlation score ≥0.80, features with higher correlation to risk labels in each anatomical direction are selected. The feature pair with the other relatively lower correlation is then removed. For feature pairs with the same correlation, those with higher variance are retained, resulting in the second set of combined features. It should be noted that this second set of combined features is obtained using risk labels as the evaluation criterion, selecting features that are more indicative of risk stratification from those with higher correlation to the risk labels.

[0142] Step 423: Sort the second combination features in descending order of importance score, and use the embedded feature selection method to filter the second combination features to obtain a subset of key features.

[0143] In some embodiments, the embedded feature selection method includes, but is not limited to, the Lasso method, the random forest method, and the LightGBM method. It should be noted that the Lasso method, the random forest method, and the LightGBM method all use risk labels as the target variable to train the model, thereby obtaining their respective importance rankings.

[0144] Specifically, the final number of features in this key feature subset is Where n is the number of features in the second combination of features, and takes the form of a positive integer. The number of low-risk examples in the training set of the machine learning model. This represents the number of high-risk examples in the training set of the machine learning model. The top 2n features are each selected using the Lasso, Random Forest, and LightGBM methods. Each of the three methods independently votes on each feature, and features receiving at least two votes are included in the candidate set. If the number of candidate examples exceeds n, the average ranking is used. (Features not selected by a certain method are penalized and ranked 2n+1) and truncated to n in ascending order; if there are fewer than n, the number is increased to 1 vote to make up the difference. ,in, The ranking of features selected using the Lasso method. This is a ranking of features selected using the Random Forest method. This is the ranking of features selected using the LightGBM method. Features receiving at least 2 votes are counted, and then sorted in descending order of importance score. Features exceeding the required number of features (n) and ranking low on the importance score are removed to ensure that the subset of key features is highly relevant to the risk stratification objective.

[0145] Step 500: Input a subset of key features of the patient to be assessed into the trained risk stratification model to perform risk stratification and output the risk level in each anatomical direction.

[0146] Specifically, the training method for this risk stratification model involves using risk labels as supervision signals and inputting a subset of key features corresponding to each historical patient into the machine learning model for training. The training steps for this machine learning model include: dividing the original dataset into training and testing sets using a stratified sampling strategy based on the risk level distribution ratio of each anatomical direction; optimizing the model hyperparameters on the training set using repeated cross-validation to obtain the optimal model hyperparameters; and retraining the optimal hyperparameters on the training set to obtain the risk stratification model and verifying its generalization performance.

[0147] It should be noted that the original dataset should be divided before feature preprocessing. The training set should be preprocessed first, and the test set should be preprocessed according to the preprocessing parameters of the training set. The purpose is to prevent the distribution of the test set from interfering with the training set and causing data leakage.

[0148] Specifically, 80% of the original dataset is allocated as the training set, and 20% is allocated as the independent test set. The training set is used for model training and hyperparameter optimization using five-fold cross-validation, while the independent test set is not involved in training and is used for independent performance evaluation. The models used for training include, but are not limited to, machine learning models such as logistic regression, support vector machines, random forests, gradient boosting trees, or deep neural networks. Preferably, the machine learning model uses the LightGBM framework, which has a hyperparameter search space of 288 combinations, including:

[0149] • The number of iterations for the tree, n_estimators: 30, 50, 100;

[0150] • Learning rate: 0.01, 0.05;

[0151] • Maximum tree depth max_depth: 3, 4;

[0152] • Minimum number of samples per leaf node: min_child_samples: 10, 20;

[0153] • L1 / L2 regularization coefficients reg_alpha / reg_lambda: 0.1, 1.0.

[0154] Specifically, K-fold cross-validation is performed on the training set, which is divided into K mutually exclusive subsets. In each iteration, one subset is selected as the validation fold and the remaining K-1 subsets are selected as the training fold. The average value of the validation fold AUC over K iterations is used as the optimization objective.

[0155] For example, five-fold cross-validation is performed on the training set, which is divided into five mutually exclusive subsets. Five rounds of training and validation are performed. In each round, one subset is selected as the validation fold and the other four subsets are selected as the training fold. That is, each round is four-fold training and one-fold validation. The average value of the AUC of the five rounds of validation folds is used as the performance index of the hyperparameter combination.

[0156] Specifically, the final risk stratification model is trained on the complete training set using the optimal hyperparameters, and the generalization performance of the risk stratification model is evaluated on an independent test set. The generalization performance metrics for evaluating the risk stratification model include, but are not limited to, AUC value, accuracy, sensitivity, and specificity.

[0157] Furthermore, the same steps are performed on the models for each anatomical direction in the original dataset. Once the models for each anatomical direction are trained, a complete risk stratification model can be obtained.

[0158] This embodiment inputs the subset of key features into a risk stratification model to stratify the risk, thereby obtaining the risk level of radiotherapy positioning uncertainty in each anatomical direction, so as to achieve personalized risk assessment before treatment.

[0159] Step 600: Based on the risk level, generate an adaptive image verification strategy and an adaptive planned target area boundary adjustment strategy.

[0160] Among them, such as Figure 11 As shown, the method for generating an adaptive image verification strategy based on risk level includes:

[0161] Step 611: Count the number of high-risk directions in each anatomical direction based on the risk level of the patient to be evaluated, and determine the initial CBCT verification frequency based on the number of high-risk directions.

[0162] Specifically, each anatomical direction includes the left-right direction, the head-to-foot direction, and the anterior-posterior direction. When the number of high-risk directions for the patient to be evaluated in each anatomical direction is 0, that is, the left-right direction, the head-to-foot direction, and the anterior-posterior direction of the patient to be evaluated are all determined to be low-risk. The initial CBCT verification frequency for the patient to be evaluated is determined to be at least one CBCT verification every 5 radiotherapy fractions, or CBCT verification is performed at least once a week as prescribed by the medical institution.

[0163] When the number of high-risk directions for a patient to be evaluated in each anatomical direction is 1, that is, any one of the left-right, head-to-foot, and front-back directions is considered high-risk and the remaining two directions are considered low-risk, the patient to be evaluated is identified as a single-direction high-risk patient to be evaluated. The initial CBCT verification frequency for the patient to be evaluated is to perform CBCT verification at least once every 4 radiotherapy fractions.

[0164] When the number of high-risk directions in each anatomical direction for the patient to be evaluated is 2, that is, any two directions of the patient to be evaluated, including the left-right direction, the head-to-foot direction, and the front-back direction, are determined to be high-risk, and the remaining direction is low-risk, the patient to be evaluated is identified as a bidirectional high-risk patient to be evaluated. The initial CBCT verification frequency for the patient to be evaluated is determined to be at least one CBCT verification every 3 radiotherapy fractions.

[0165] When the number of high-risk directions in each anatomical direction for the patient to be evaluated is 3, that is, the left-right direction, head-to-foot direction and front-back direction of the patient to be evaluated are all determined to be high-risk, the patient to be evaluated is identified as a three-direction high-risk patient to be evaluated, and the initial CBCT verification frequency for the patient to be evaluated is determined to be at least one CBCT verification every 2 radiotherapy fractions.

[0166] Step 612: Dynamically adjust the initial CBCT verification frequency based on the radiotherapy positioning error data obtained from online CBCT to obtain the second CBCT verification frequency.

[0167] The dynamic adjustment includes: acquiring radiotherapy positioning error data for each radiotherapy fraction, and calculating the absolute value of the radiotherapy positioning error data for each anatomical direction; monitoring the number of times the radiotherapy positioning error data for each anatomical direction exceeds the corresponding preset error threshold within a preset observation window; when the number of times the radiotherapy positioning error data for at least one anatomical direction exceeds the corresponding preset error threshold reaches a preset threshold, it is determined that the patient to be evaluated has a persistent positioning risk in the corresponding anatomical direction, and the initial CBCT verification frequency is increased to a second CBCT verification frequency.

[0168] Specifically, the second CBCT verification frequency involves performing one CBCT verification for each radiotherapy fraction.

[0169] Step 613: When the online radiotherapy positioning error data of each anatomical direction of the patient to be evaluated is lower than the corresponding recovery threshold in a consecutive preset number of radiotherapy fractions, the risk level is updated according to the radiotherapy positioning error data of the existing CBCT fractions in the radiotherapy cycle, and the second CBCT verification frequency is restored to the CBCT verification frequency corresponding to the updated risk level.

[0170] The criteria for determining the risk level mentioned above can be found in steps 412 and 413.

[0171] Preferably, the preset observation window, preset error threshold, preset number of times threshold, recovery threshold, and consecutive preset number of times are as follows: the preset observation window is 5 consecutive radiotherapy fractions that have been verified by CBCT; the preset error threshold is set to 0.3 cm in the left-right direction, 0.5 cm in the head-to-toe direction, and 0.3 cm in the front-back direction; the preset number of times threshold is at least 3 times within the preset observation window; the recovery threshold is set to 0.3 cm in the left-right direction, 0.5 cm in the head-to-toe direction, and 0.3 cm in the front-back direction; and the consecutive preset number of times of radiotherapy fractions is preferably 3 consecutive radiotherapy fractions that have been verified by CBCT.

[0172] It should be noted that the parameters of every 5 radiotherapy fractions, every 4 radiotherapy fractions, every 3 radiotherapy fractions, every 2 radiotherapy fractions, each radiotherapy fraction, as well as the preset observation window, preset error threshold, preset number of times threshold, recovery threshold, and consecutive preset number are only illustrative parameters. The actual values ​​can be set or adjusted according to at least one of the following: the risk level of the patient to be evaluated, the treatment site, the total number of treatment fractions, the type of fixation device, the distribution of historical radiotherapy positioning error data, the medical institution's imaging verification standards, and the confirmation results of the radiotherapy physicist or physician.

[0173] In this embodiment, the risk levels of three anatomical directions output by the risk stratification model are used to determine the CBCT verification frequency: when the number of high-risk directions increases, the initial CBCT verification frequency is increased; when the number of times the online radiotherapy positioning error data in at least one anatomical direction exceeds the corresponding preset error threshold during treatment reaches the preset number threshold, the CBCT verification frequency is further increased; when the online radiotherapy positioning error data remains below the recovery threshold for multiple consecutive treatment fractions, the CBCT verification frequency is reduced or restored, thereby forming an adaptive image verification process based on pre-treatment risk prediction and in-treatment error feedback.

[0174] like Figure 12 As shown, the method for generating an adaptive target area boundary adjustment strategy based on risk level includes:

[0175] Step 621: Divide each historical patient into a low-risk group and a high-risk group based on the risk labels of each historical patient in each anatomical direction.

[0176] Specifically, each anatomical direction includes the left-right direction, the head-to-foot direction, and the anterior-posterior direction. For ease of description, the left-right direction is denoted as the x-direction, the head-to-foot direction as the y-direction, and the anterior-posterior direction as the z-direction. Each anatomical direction is uniformly denoted as d, i.e., d∈{x,y,z}. In the training set of the machine learning model, for each historical patient, risk labels are obtained in the left-right, head-to-foot, and anterior-posterior directions. These risk labels characterize the patient's tendency to have a large positioning error in the corresponding anatomical direction, and their values ​​include low risk and high risk. Based on the risk labels in the left-right, head-to-foot, and anterior-posterior directions, historical patients are divided into a low-risk group and a high-risk group. The low-risk group consists of patients who are classified as low-risk in all three directions; the high-risk group consists of patients who are classified as high-risk in at least one of these three directions. For ease of description, the low-risk group is designated as the first group, i.e., g=1; the high-risk group is designated as the second group, i.e., g=2; and any group is uniformly designated as the g-th group, g∈{1,2}.

[0177] Step 622: Obtain the first historical radiotherapy positioning error data of historical patients in the low-risk group in each anatomical direction, and determine the first systematic error data and the first random error data of historical patients in the low-risk group in each anatomical direction based on the first historical radiotherapy positioning error data; obtain the second historical radiotherapy positioning error data of historical patients in the high-risk group in each anatomical direction, and determine the second systematic error data and the second random error data of historical patients in the high-risk group in each anatomical direction based on the second historical radiotherapy positioning error data.

[0178] Specifically, the first systematic error data includes a first population systematic error, and the first random error data includes a first population random error; the second systematic error data includes a second population systematic error, and the second random error data includes a second population random error. To obtain the aforementioned population-level error data, it is necessary to first determine the error data at the individual level for each historical patient, and then aggregate the individual-level error data to obtain the population-level error data. The specific process is as follows:

[0179] Let the g-th group contain N. g There are n historical patients in this group. The i-th historical patient in this group is denoted as historical patient (g, i). This historical patient has undergone n anatomical procedures along the anatomical direction d. g,i The positional error is measured in the j-th measurement, and the positional error obtained from the j-th measurement is denoted as Δ. g,i,d,j .

[0180] (a) Individual systematic error. For a historical patient (g, i) in group g, the individual systematic error m in direction d is...g,i,d m is the average value of the positioning error obtained from all measurements in direction d for this historical patient. g,i,d The specific calculation method is as follows .

[0181] (ii) Group systematic error. The group systematic error Σ of group g in direction d. g,d (When g=1, this represents the systematic error of the first group; when g=2, this represents the systematic error of the second group.) The individual systematic error m represents all historical patients within this group. g,i,d Standard deviation, Σ g,d The specific calculation formula is as follows , where μ g,d Let be the average of the individual systematic errors of all historical patients within group g. .

[0182] (iii) Individual random error. For a historical patient (g, i) in group g, the individual random error σ in direction d is... g,i,d The positioning error obtained from each measurement in direction d for this historical patient is relative to its individual systematic error m. g,i,d Standard deviation, σ g,i,d The specific calculation method is as follows .

[0183] (iv) Population random error. The population random error σ of the g-th group in direction d. g,d (When g=1, this represents the random error of the first group; when g=2, this represents the random error of the second group.) σ represents the individual random error of all historical patients within this group. g,i,d The root mean square of σ g,d The specific calculation method is as follows .

[0184] Therefore, setting g=1, we obtain the first group systematic error Σ for the low-risk group in each anatomical direction. 1,d With the first group random error σ 1,d Let g=2, then we obtain the second group systematic error Σ for this high-risk group in each anatomical direction. 2,d With the second group random error σ 2,d , where d∈{x,y,z}.

[0185] It should be noted that the first and second group systematic errors represent recurring deviations in the radiotherapy process that are consistent in direction and cannot be canceled out by multiple fractionated treatments. Their impact on target dose coverage needs to be fully covered by the outer boundary. The first and second group random errors represent deviations caused by random fluctuations in each fractionated treatment during the radiotherapy process, which can be partially canceled out by multiple fractionated treatments. Therefore, in the subsequent calculation of the outer boundary, systematic errors (first and second group systematic errors) and random errors (first and second group random errors) are assigned different weighting coefficients.

[0186] Step 623: Based on the first systematic error data and the first random error data, calculate the first planned target area expansion boundary of the low-risk group in each anatomical direction using the van Herk formula; based on the second systematic error data and the second random error data, calculate the second planned target area expansion boundary of the high-risk group in each anatomical direction using the van Herk formula.

[0187] Specifically, the van Herk formula was used to calculate the risk level of this low-risk group in any anatomical direction. The first planned target area expansion boundary The specific calculation method is as follows , where d∈{x,y,z}, Σ 1,d For the low-risk group, the first group systematic error in direction d is σ. 1,d The coefficient 2.5 represents the first group random error in direction d for the low-risk group. It is used to ensure that at least 90% of patients in the low-risk group can obtain no less than 95% of the prescribed dose in their clinical target volume (CTV), thereby providing sufficient coverage for uncompensated systematic errors. The coefficient 0.7 represents the random error (first group random error and second group random error) and is used to compensate for the broadening effect of random errors on the dose distribution edge (penumbra).

[0188] Similarly, the van Herk formula is used to calculate the outer boundary M of the second planned target area for this high-risk group in any anatomical direction d. 2,d The specific calculation method is as follows , where d∈{x,y,z}; Σ 2,d For the high-risk group, the second group systematic error in direction d is σ. 2,dThe second group random error in direction d represents the high-risk group. The coefficient 2.5 is the coefficient of the systematic error term (first group systematic error and second group systematic error), which is used to ensure that at least 90% of patients in the high-risk group can obtain no less than 95% of the prescribed dose in their clinical target volume (CTV), thereby providing sufficient coverage for the uncompensated systematic error. The coefficient 0.7 is the coefficient of the random error term (first group random error and second group random error), which is used to compensate for the broadening effect of random error on the dose distribution edge (penumbra).

[0189] Therefore, the outer boundaries of the first planned target area for the low-risk group in the left-right, head-to-toe, and front-to-back directions are M respectively. 1,x M 1,y M 1,z The high-risk group's second planned target area expansion boundaries in the left-right, head-to-toe, and front-to-back directions are M respectively. 2,x M 2,y M 2,z .

[0190] Step 624: Based on the risk level of each patient to be evaluated in each anatomical direction, divide each patient to be evaluated into a low-risk group and a high-risk group. For the patients to be evaluated in the low-risk group, expand the clinical target area of ​​the patient to be evaluated in the corresponding anatomical direction based on the outer boundary of the first planning target area to obtain the first planning target area. For the patients to be evaluated in the high-risk group, expand the clinical target area of ​​the patient to be evaluated in the corresponding anatomical direction based on the outer boundary of the second planning target area to obtain the second planning target area.

[0191] Specifically, in this embodiment, the criteria for classifying patients to be evaluated into low-risk and high-risk groups based on their risk levels in each anatomical direction are similar to those described above for classifying historical patients into low-risk and high-risk groups based on their risk labels in each anatomical direction, and will not be elaborated upon further here. For patients in the low-risk group, the criteria for classifying them into low-risk and high-risk groups in the left-right, head-to-foot, and anterior-posterior directions of their clinical target area are respectively based on M... 1,x M 1,y M 1,z Anisotropic expansion was performed to obtain the planned target area (i.e., the first planned target area) for patients to be evaluated in the low-risk group; for patients to be evaluated in the high-risk group, the clinical target area was expanded according to M in the left-right, head-to-foot, and anteroposterior directions. 2,x M 2,y M 2,z Anisotropic expansion was performed to obtain the planned target area (i.e., the second planned target area) for patients to be evaluated in this high-risk group.

[0192] Both the first and second planned target areas are not lower than the minimum clinical safety threshold to ensure that the clinical target area coverage meets the prescription requirements.

[0193] In some embodiments, the minimum clinical safety threshold can be determined based on at least one of the following: treatment site, target type, treatment technique, image-guided method, and radiotherapy quality control standards of the medical institution.

[0194] It should be noted that this minimum clinical safety threshold is a threshold preset by the medical institution and verified by a radiation therapy physicist or physician.

[0195] This embodiment utilizes differentiated CBCT verification frequencies and radiotherapy target zone boundary adjustments. Specifically, for patients in the high-risk group, the CBCT verification frequency is increased, and the outer boundary of the radiotherapy target zone is appropriately expanded to compensate for the greater uncertainty in radiotherapy positioning. For patients in the low-risk group, the outer boundary of the radiotherapy target zone is appropriately reduced to decrease the radiation dose to surrounding normal tissues. While ensuring the accuracy of radiotherapy treatment, this helps reduce unnecessary additional radiation exposure for patients and optimizes the allocation of radiotherapy resources.

[0196] In another exemplary embodiment, such as Figure 13 As shown, this application also provides a method for risk stratification and adaptive decision-making regarding radiotherapy positioning uncertainty, the method comprising:

[0197] Step 1101: Acquire multimodal data of the patient, including target CT images and three-dimensional body surface data;

[0198] In this embodiment, step 1101 is similar to step 100 above, and will not be described again here.

[0199] Step 1102: Obtain the spatial mapping relationship between the three-dimensional body surface data and the target CT image, and perform spatial alignment of the three-dimensional body surface data based on the spatial mapping relationship to obtain the aligned three-dimensional body surface data.

[0200] In this embodiment, the three-dimensional body surface data is acquired by a body surface acquisition device (such as an SGRT device) and is located in the coordinate system of the body surface acquisition device itself, which differs from the coordinate system of the CT image. Therefore, it is necessary to spatially align the three-dimensional body surface data with the target CT image. This spatial alignment is performed by using a spatial affine transformation matrix or a rigid registration algorithm based on common feature points to transform the three-dimensional body surface data to the coordinate system space of the target CT image.

[0201] Specifically, firstly, a threshold segmentation algorithm or edge detection algorithm is used to extract the patient's body contour from the target CT image to generate a reference body surface model, which is then used as the registration benchmark. Next, the three-dimensional body surface data acquired by the body surface acquisition device is registered and transformed to the coordinate system space of the target CT image where the reference body surface model is located using a spatial affine transformation matrix or a rigid registration algorithm based on common feature points. The three-dimensional body surface data acquired by the body surface acquisition device after registration and transformation is the aligned three-dimensional body surface data, which is used for subsequent feature extraction.

[0202] The rigid registration algorithm is as follows: the three-dimensional body surface data acquired by the body surface acquisition device is recorded in the form of point cloud or mesh to record the spatial coordinates of the patient's body surface; coarse registration is performed using body surface markers or preset mechanical coordinate transformation parameters of the body surface acquisition device to initially align the body surface point cloud to the approximate position of the reference body surface model; then, in the fine registration stage, the correspondence between the nearest points of the body surface point cloud and the surface of the reference body surface model is calculated in each iteration, the overall mean square distance error between the corresponding point sets is evaluated, and the optimal rotation that minimizes the overall mean square distance error is solved through matrix operations. The optimal rotation matrix and translation vector are applied to the surface point cloud. The nearest point correspondence is searched again and the overall mean square distance error is calculated. This process is iterated until the error change between two consecutive iterations is less than a preset threshold or the maximum number of iterations is reached. At this point, convergence is determined, and the optimal spatial transformation matrix is ​​obtained. After obtaining the optimal spatial transformation matrix, it is applied to the preprocessed 3D surface data. The local coordinates of all vertices in the surface point cloud are transformed to the coordinate system of the target CT image through matrix multiplication to generate aligned 3D surface data.

[0203] In some embodiments, the preset threshold can be set by those skilled in the art based on historical experimental experience.

[0204] It should be noted that the three-dimensional body surface data acquired by the body surface acquisition device after registration transformation using a rigid registration algorithm, rather than the reference body surface model reconstructed from the target CT image, is used for subsequent feature extraction. This is because the three-dimensional body surface data acquired by the acquisition device can reflect the patient's true body surface morphology (such as surface undulations, local occlusions, etc.) in a state closer to the actual treatment positioning, and carries independent body surface information that the target CT image cannot provide. At this time, the reference body surface model reconstructed from the target CT image is only used as a registration benchmark to transform the externally acquired data to the CT image coordinate system.

[0205] This embodiment, by acquiring spatial mapping relationships and spatially aligning the three-dimensional body surface data, can eliminate the coordinate system differences between the body surface data and the CT equipment when the data is acquired by external devices. When the body surface data is reconstructed by CT, its inherent coordinate system can be directly reused, and the three-dimensional body surface data can be uniformly mapped to the CT image coordinate system of the target CT image. This ensures that any point on the patient's body surface corresponds one-to-one with a voxel in the target CT image in space, thereby ensuring that the features extracted from the three-dimensional body surface data and the target CT image have strict spatial consistency in anatomical location, laying a spatial benchmark for subsequent feature extraction and risk level prediction.

[0206] Step 1103: Use a deep learning model to segment the target CT image to obtain body component labels and anatomical markers.

[0207] Step 1104: Perform bed removal processing on the aligned three-dimensional body surface data and combine it with the anatomical markers to determine the region of interest on the patient's body surface.

[0208] Step 1105: Extract the first type of features from the region of interest on the patient's body surface and the second type of features from the target CT image. Perform feature processing on the first type of features, the second type of features, and clinical medical features to obtain a subset of key features.

[0209] Step 1106: Input the subset of key features of the patient to be evaluated into the trained risk stratification model to perform risk stratification and output the risk level in each anatomical direction.

[0210] Step 1107: Based on the risk level, generate an adaptive image verification strategy and an adaptive planned target area boundary adjustment strategy.

[0211] In this embodiment, steps 1103 to 1107 are similar to steps 200 to 600 described above, and will not be repeated here.

[0212] It should be noted that, without any limitation, the patients mentioned above in this application refer to patients to be evaluated and historical patients.

[0213] The scope of protection of the radiotherapy position uncertainty risk stratification and adaptive decision-making method described in this embodiment is not limited to the order of steps listed in this embodiment. Any solution implemented by adding, subtracting or replacing steps in the prior art based on the principle of this invention is included within the scope of protection of this invention.

[0214] This invention also provides a radiotherapy position uncertainty risk stratification and adaptive decision-making system. The radiotherapy position uncertainty risk stratification and adaptive decision-making system can implement the radiotherapy position uncertainty risk stratification and adaptive decision-making method described in this invention. However, the implementation device of the radiotherapy position uncertainty risk stratification and adaptive decision-making method described in this invention includes, but is not limited to, the structure of the radiotherapy position uncertainty risk stratification and adaptive decision-making system listed in this embodiment. All structural modifications and substitutions of the prior art made according to the principles of this invention are included within the protection scope of this invention.

[0215] like Figure 14 As shown, this embodiment provides a radiotherapy positioning uncertainty risk stratification and adaptive decision-making system. The system includes: a data acquisition module 1, configured to acquire multimodal data of the patient, including target CT images and three-dimensional body surface data; a data processing module 2, configured to use a deep learning model to segment the target CT images to obtain body composition labels and anatomical markers; a region of interest generation module 3, configured to perform debedming on the three-dimensional body surface data and, in conjunction with the anatomical markers, determine the region of interest on the patient's body surface; a feature processing module 4, configured to extract a first type of feature from the region of interest on the patient's body surface and a second type of feature from the target CT images, and perform feature processing on the first type of feature, the second type of feature, and clinical medical features to obtain a key feature subset; a risk prediction module 5, configured to input the key feature subset of the patient to be evaluated into a trained risk stratification model for risk stratification and output the risk level in each anatomical direction; and an adaptive decision-making module 6, configured to generate an adaptive image verification strategy and an adaptive planning target boundary adjustment strategy based on the risk level.

[0216] In some embodiments, the data processing module 2 is further configured to acquire the spatial mapping relationship between the three-dimensional body surface data and the target CT image, and to spatially align the three-dimensional body surface data based on the spatial mapping relationship to obtain aligned three-dimensional body surface data.

[0217] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, or methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or units may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or modules or units may be electrical, mechanical, or other forms.

[0218] The modules / units described as separate components may or may not be physically separate. The components shown as modules / units may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules / units can be selected to achieve the objectives of the embodiments of the present invention, depending on actual needs. For example, the functional modules / units in the various embodiments of the present invention may be integrated into one processing module, or each module / unit may exist physically separately, or two or more modules / units may be integrated into one module / unit.

[0219] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0220] In one exemplary embodiment, this embodiment also provides an electronic device including a memory and a processor, wherein the memory stores a program executable on the processor, and when the program is executed by the processor, the electronic device performs any of the methods described in the above embodiments.

[0221] In one possible embodiment, such as Figure 15 As shown, the electronic device 20 also includes: an output interface 23 for outputting results; a communication interface 24 for transmitting communication signals; and an antenna 25 for transmitting or receiving signals.

[0222] It should be noted that the processor 21 in this embodiment can be an image processing chip or an integrated circuit chip, capable of processing image signals. In implementation, each step of the above method embodiment can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can 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. It can implement or execute the methods, steps, and logic block diagrams disclosed in this embodiment. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this embodiment can be directly implemented by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0223] Those skilled in the art will understand that Figure 15 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the electronic device to which the present invention is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0224] In one exemplary embodiment, this embodiment also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps in the above-described method embodiments.

[0225] In one exemplary embodiment, this embodiment also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0226] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations.

[0227] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided by this invention may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided by this invention may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0228] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A risk stratification and adaptive decision-making method for radiotherapy positioning uncertainty, characterized in that, The method includes: Acquire multimodal data of the patient, including target CT images and three-dimensional body surface data; A deep learning model is used to segment the target CT image to obtain body component labels and anatomical markers; The three-dimensional body surface data is processed to remove bed material, and combined with the anatomical markers, the region of interest on the patient's body surface is determined. The first type of features are extracted from the region of interest on the patient's body surface, and the second type of features are extracted from the target CT image. The first type of features, the second type of features, and the clinical medical features are then processed to obtain a subset of key features. A subset of key features of the patient to be assessed is input into a trained risk stratification model to perform risk stratification and output the risk level in each anatomical direction. Based on the risk level, an adaptive image verification strategy and an adaptive planned target area boundary adjustment strategy are generated.

2. The method according to claim 1, characterized in that, The process of removing bed material from the three-dimensional body surface data and combining it with the anatomical markers to determine the region of interest on the patient's body surface includes: Spatial morphological processing is performed on the three-dimensional body surface data to obtain primary body surface data; A back reference plane is established based on the anatomical features of the primary body surface data, and interfering data below the back reference plane is removed to obtain the corrected body surface data. The corrected body surface data is extracted along the long axis of the human body to obtain the purified body surface data; Based on the purified body surface data and the anatomical markers, the region of interest on the patient's body surface is determined.

3. The method according to claim 1, characterized in that, Before performing feature processing on the first type of features, the second type of features, and the clinical medical features, the following steps are included: Historical radiotherapy positioning error data for each patient was obtained, including positioning error measurements in each anatomical direction. Based on the historical radiotherapy positioning error data, the mean and standard deviation of the absolute values ​​of positioning errors in each anatomical direction for each historical patient were calculated. Based on the mean, the standard deviation, and the preset risk assessment threshold, risk labels for each anatomical direction of each historical patient are generated.

4. The method according to claim 3, characterized in that, The process involves generating risk labels for each historical patient in each anatomical direction based on the mean, the standard deviation, and a preset risk assessment threshold. These anatomical directions include left-right, head-to-foot, and anteroposterior directions. For the left-right and front-back directions, a first judgment threshold is calculated based on the mean and standard deviation. If the first judgment threshold is greater than the first preset risk judgment threshold, the corresponding direction is judged as high risk; otherwise, it is judged as low risk. For the head-to-toe direction, a second judgment threshold is calculated based on the mean and standard deviation. If the second judgment threshold is greater than the second preset risk judgment threshold, the head-to-toe direction is judged as high risk; otherwise, it is judged as low risk.

5. The method according to claim 3, characterized in that, The training method for the risk stratification model includes: Using the risk labels as supervisory signals, the key feature subsets corresponding to each historical patient are input into a machine learning method for training to obtain the risk stratification model.

6. The method according to claim 3, characterized in that, The process of generating an adaptive image verification strategy and an adaptive target area boundary adjustment strategy based on the risk level includes: The adaptive image verification strategy includes: counting the number of high-risk directions in each anatomical direction based on the risk level of the patient to be evaluated in each anatomical direction, and determining the initial CBCT verification frequency based on the number of high-risk directions; The initial CBCT verification frequency is dynamically adjusted based on the radiotherapy positioning error data obtained from online CBCT. This dynamic adjustment includes: acquiring radiotherapy positioning error data for each treatment fraction and calculating the absolute value of the radiotherapy positioning error data for each anatomical direction; monitoring the number of times the radiotherapy positioning error data for each anatomical direction exceeds a corresponding preset error threshold within a preset observation window; when the number of times the radiotherapy positioning error data for at least one anatomical direction exceeds the corresponding preset error threshold reaches a preset threshold, it is determined that the patient to be evaluated has a persistent positioning risk in the corresponding anatomical direction, and the initial CBCT verification frequency is increased to a second CBCT verification frequency. When the online radiotherapy positioning error data of each anatomical direction of the patient to be evaluated is lower than the corresponding recovery threshold in a continuous preset number of radiotherapy fractions, the risk level is updated according to the radiotherapy positioning error data of the existing CBCT fractions in the radiotherapy cycle, and the second CBCT verification frequency is restored to the CBCT verification frequency corresponding to the updated risk level. The adaptive planning target boundary adjustment strategy includes: classifying each historical patient into a low-risk group and a high-risk group based on the risk labels of each historical patient in each anatomical direction; Acquire first historical radiotherapy positioning error data for historical patients in the low-risk group in each anatomical direction, and determine first systematic error data and first random error data for historical patients in the low-risk group in each anatomical direction based on the first historical radiotherapy positioning error data; acquire second historical radiotherapy positioning error data for historical patients in the high-risk group in each anatomical direction, and determine second systematic error data and second random error data for historical patients in the high-risk group in each anatomical direction based on the second historical radiotherapy positioning error data; Based on the first systematic error data and the first random error data, the van Herk formula is used to calculate the first planned target area expansion boundary of the low-risk group in each anatomical direction; based on the second systematic error data and the second random error data, the van Herk formula is used to calculate the second planned target area expansion boundary of the high-risk group in each anatomical direction. Based on the risk level of each patient under evaluation in each anatomical direction, the patients under evaluation are divided into a low-risk group and a high-risk group. For the patients under evaluation in the low-risk group, the clinical target area of ​​the patients under evaluation is expanded outward in the corresponding anatomical direction based on the outer boundary of the first planned target area to obtain the first planned target area. For the patients under evaluation in the high-risk group, the clinical target area of ​​the patients under evaluation is expanded outward in the corresponding anatomical direction based on the outer boundary of the second planned target area to obtain the second planned target area. Both the first planned target area and the second planned target area are not lower than the minimum clinical safety threshold.

7. The method according to claim 1, characterized in that, Before performing feature processing on the first type of features, the second type of features, and the clinical medical features, the method further includes: The continuous features in the first type of features, the second type of features, and clinical medical features are truncated at 1% to 99% quantiles, and feature values ​​that exceed the 1% to 99% quantile range are replaced with corresponding boundary feature values. The first type of features, the second type of features, and the multi-category nominal features in clinical medical features are converted into 0 or 1 integer vectors using one-hot encoding. The first type of features, the second type of features, and clinical medical features are standardized. The first type of features includes at least multidimensional body surface morphological features, and the second type of features includes at least radiomics features, spatial anatomical relationship features, and body composition quantification features.

8. The method according to claim 7, characterized in that, Feature processing is performed on the first type of features, the second type of features, and clinical medical features to obtain a subset of key features, including: Features whose variance is less than a first preset threshold among the first type of features, the second type of features, and clinical medical features are removed to obtain the first combined features; The first combined features are filtered using the Pearson correlation coefficient to obtain the second combined features; The second combined features are sorted in descending order of importance score, and the embedded feature selection method is used to filter the second combined features to obtain a subset of key features.

9. The method according to claim 7, characterized in that, The anatomical markers include a first anatomical marker, a second anatomical marker, and a third anatomical marker. The calculation method for the body composition quantification characteristics is as follows: Using the cardiac level of the first anatomical marker material as the upper boundary and the cardiac level of the second anatomical marker material as the lower boundary, the target CT image is cropped to determine the range of three-dimensional volume composition analysis. The cardiac level, the third anatomical marker, is used as the scope for two-dimensional body composition analysis. The volume composition quantification features are obtained based on the two-dimensional volume composition analysis range and the three-dimensional volume composition analysis range.

10. The method according to claim 7, characterized in that, The multidimensional surface morphological features include a group of micro-statistical features reflecting the statistical distribution of point cloud shape index and curvature, a group of global geometric features reflecting the overall geometric shape, a group of local structural features reflecting the distribution of local curvature and normal, and a group of topological features reflecting the integrity of the mesh topology. The image omics features include geometric shape descriptors, voxel gray value statistics, and texture descriptors; The spatial anatomical relationship features include: the minimum three-dimensional Euclidean distance from the centroid of the target radiotherapy area to the body surface; the minimum three-dimensional distance from the centroid of the target radiotherapy area to the bone; the minimum three-dimensional Euclidean distance from the isocenter to the body surface; the minimum three-dimensional distance from the isocenter to the bone; the three-dimensional Euclidean distance from the isocenter to the centroid of the target radiotherapy area and the components of the three-dimensional Euclidean distance in each direction; and, on the cross-section passing through the isocenter, the ratio of the anteroposterior diameter to the left-right diameter and the perimeter of the body surface contour at the isocenter level.

11. The method according to claim 1, characterized in that, Before segmenting the target CT image using a deep learning model, the process includes: The spatial mapping relationship between the three-dimensional body surface data and the target CT image is obtained, and the three-dimensional body surface data is spatially aligned based on the spatial mapping relationship to obtain aligned three-dimensional body surface data.

12. The method according to claim 1, characterized in that, The body composition label includes at least subcutaneous fat, visceral fat, muscle tissue, intramuscular fat, and bone tissue.

13. A risk stratification and adaptive decision-making system for radiotherapy positioning uncertainty, characterized in that, The system includes: The data acquisition module is configured to acquire multimodal data of the patient, including target CT images and three-dimensional body surface data; The data processing module is configured to use a deep learning model to segment the target CT image to obtain body component labels and anatomical markers; The region of interest generation module is configured to perform bed removal processing on the three-dimensional body surface data and, in conjunction with the anatomical markers, determine the region of interest on the patient's body surface. The feature processing module is configured to extract a first type of feature from the region of interest on the patient's body surface, extract a second type of feature from the target CT image, and perform feature processing on the first type of feature, the second type of feature, and clinical medical features to obtain a subset of key features. The risk prediction module is configured to input a subset of key features of the patient to be assessed into a trained risk stratification model to perform risk stratification and output the risk level in each anatomical direction. The adaptive decision-making module is configured to generate an adaptive image verification strategy and an adaptive target area boundary adjustment strategy based on the risk level.

Images