Chest tumor intensity modulated radiotherapy aided analysis method based on AI dose prediction

Abstract

The invention relates to the technical field of radiotherapy plan assistance, and discloses a chest tumor intensity modulated radiotherapy assisted analysis method based on AI dose prediction, and the method comprises the steps: obtaining a four-dimensional CT, a plan CT, a radiotherapy plan, and an AI initial dose prediction result; establishing a treatment coordinate system and unifying coordinates; calculating a three-dimensional electron density gradient field and a cleaned three-dimensional deformation vector field; analyzing the beam to generate a two-dimensional flux gradient, and calculating a physical instability index in combination with a projection motion vector; calculating an anatomical motion coupling index based on the deformation field and the density gradient field; and fusing the two indexes to generate a failure risk index and a confidence map, and performing superposition display in a thermodynamic diagram form. According to the method, a high-risk area of AI prediction failure is visually recognized by quantifying the flux cutting and anatomical motion coupling effect, the defect that a static model neglects dynamic characteristics is overcome, and the reliability of chest tumor intensity modulated radiation therapy auxiliary analysis is improved.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy planning assistance technology, specifically to an AI-based dose prediction-assisted analysis method for intensity-modulated radiotherapy (IMRT) of thoracic tumors. Background Technology

[0002] Radiotherapy for thoracic tumors is challenging due to the large amplitude of respiratory movements and the high heterogeneity of tissue density. Intensity-modulated radiotherapy (IMRT), as the mainstream technology for treating thoracic tumors, uses multi-leaf gratings to finely modulate the beam flux intensity to achieve high dose coverage of the target area and effective protection of surrounding organs at risk. Its treatment process involves complex physical modulation and the interaction of anatomical structures.

[0003] Currently, with the development of artificial intelligence technology, deep learning-based AI dose prediction technology has been widely applied in intensity-modulated radiotherapy (IMRT) auxiliary analysis. This type of technology typically utilizes convolutional neural networks to learn the characteristics of historical high-quality treatment plans, automatically generating predicted three-dimensional dose distributions based on the patient's anatomical images and delineated contours. It aims to improve the efficiency and consistency of treatment plan design and assist physicists in evaluating plan quality.

[0004] However, most existing AI dose prediction models are trained and inferred based on static anatomical images, often neglecting the dynamic characteristics of thoracic tumors during actual treatment delivery. Current technologies struggle to effectively capture the flux-cutting effect resulting from the interaction between the dynamic modulation of multi-leaf gratings and the patient's respiratory movements, and also fail to quantify the specific impact of drastic changes in the water equivalent depth caused by tissue movement along the electron density gradient on dose deposition accuracy. This neglect of dynamic physical mechanisms and anatomical motion coupling characteristics means that while AI models can output dose distributions when dealing with highly dynamic and complex cases, they cannot identify potential prediction failure risks due to motion and modulation conflicts, limiting the reliability and safety of AI dose prediction in clinical intensity-modulated radiotherapy (IMRT)-assisted analysis.

[0005] Therefore, this invention proposes an AI-based dose prediction-assisted analysis method for intensity-modulated radiotherapy of thoracic tumors to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an AI-based dose prediction-assisted analysis method for intensity-modulated radiotherapy (IMRT) of thoracic tumors. This method solves the problem that existing AI dose prediction models based on static images are unable to perceive the dose deviation risk caused by the interaction between respiratory motion, beam modulation, and tissue density changes during IMRT of thoracic tumors, thus making it difficult to assess the reliability of the prediction results.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an AI-based dose prediction-assisted analysis method for intensity-modulated radiotherapy (IMRT) of thoracic tumors, comprising the following steps: S110: Obtain the four-dimensional CT image sequence, planned CT image, intensity-modulated radiotherapy planning file, and initial dose prediction results generated by the artificial intelligence model for the case to be analyzed; S120, parse the isocenter position information in the intensity-modulated radiotherapy planning file, establish a treatment coordinate system with the isocenter as the origin, and uniformly convert the voxel coordinates of the four-dimensional CT image sequence and the planning CT image to the treatment coordinate system; S130, calculate the three-dimensional electron density gradient field based on the planned CT image, and calculate the three-dimensional deformation vector field based on the four-dimensional CT image sequence, and perform consistency cleaning and regularization processing on the three-dimensional deformation vector field; S140, parse the beam parameters in the intensity-modulated radiotherapy planning file, generate a two-dimensional flux map corresponding to the beam viewpoint, and calculate the two-dimensional flux gradient; construct the projection relationship of the three-dimensional deformation vector field on the beam plane according to the treatment coordinate system, and calculate the physical instability index in combination with the two-dimensional flux gradient; S150, Calculate the anatomical motion coupling index based on the three-dimensional deformation vector field and the three-dimensional electron density gradient field; S160, by fusing the physical instability index and the anatomical motion coupling index, a predicted failure risk index is generated, and a confidence map of the AI ​​dose prediction is determined. The confidence map is then overlaid on the planned CT image and the initial dose prediction result in the form of a heatmap.

[0008] Preferably, in step S130, the step of calculating the three-dimensional electron density gradient field based on the planned CT image includes: By using a pre-configured CT value to electron density conversion protocol, the gray value of each voxel in the planned CT image is converted into an electron density value to generate a three-dimensional electron density volume. The three-dimensional electron density gradient field is obtained by applying a three-dimensional difference operator or a Sobel operator to the three-dimensional electron density volume to calculate the spatial gradient vector of each voxel point.

[0009] Preferably, in step S130, the step of calculating the three-dimensional deformation vector field based on the four-dimensional CT image sequence includes: The end-expiratory phase image in the four-dimensional CT image sequence is selected as the source image, and the end-inspiratory phase image is selected as the target image. A non-rigid registration operation is performed on the source image and the target image to calculate the displacement vector of each voxel in the source image moving to the corresponding position in the target image, thereby generating the original three-dimensional deformation vector field.

[0010] Preferably, in step S130, the step of calculating the three-dimensional deformation vector field based on the four-dimensional CT image sequence includes: The end-expiratory phase image in the four-dimensional CT image sequence is selected as the source image, and the end-inspiratory phase image is selected as the target image. A non-rigid registration operation is performed on the source image and the target image to calculate the displacement vector of each voxel in the source image moving to the corresponding position in the target image, thereby generating the original three-dimensional deformation vector field.

[0011] Preferably, in step S140, the step of parsing the beam parameters in the intensity-modulated radiotherapy (IMRT) planning file, generating a two-dimensional flux map of the corresponding beam angle, and calculating the two-dimensional flux gradient includes: Traverse each beam in the intensity-modulated radiotherapy planning file and construct a two-dimensional grid plane perpendicular to the beam's central axis; Based on the control point sequence of the multi-leaf grating blade positions and the corresponding subfield weights, the cumulative radiative flux intensity of each pixel on the two-dimensional grid plane is calculated to generate the two-dimensional flux map. Perform convolution calculations on the two-dimensional flux graph to obtain the two-dimensional flux gradient.

[0012] Preferably, in step S140, the step of constructing the projection relationship of the three-dimensional deformation vector field onto the beam plane based on the treatment coordinate system, and calculating the physical instability index in conjunction with the two-dimensional flux gradient includes: Based on the frame angle and collimator angle of the beam, a three-dimensional to two-dimensional projection matrix is ​​constructed; The three-dimensional deformation vector field is rotated to the beam viewpoint using the projection matrix, and the depth motion component along the beam axis is removed to obtain the projected motion vector; Calculate the dot product of the projected motion vector and the two-dimensional flux gradient, and take the absolute value of the dot product; The physical instability index is obtained by weighted summation of the absolute values ​​calculated for all beams.

[0013] Preferably, in step S140, the step of constructing the projection relationship of the three-dimensional deformation vector field onto the beam plane based on the treatment coordinate system, and calculating the physical instability index in conjunction with the two-dimensional flux gradient includes: Based on the frame angle and collimator angle of the beam, a three-dimensional to two-dimensional projection matrix is ​​constructed; The three-dimensional deformation vector field is rotated to the beam viewpoint using the projection matrix, and the depth motion component along the beam axis is removed to obtain the projected motion vector; Calculate the dot product of the projected motion vector and the two-dimensional flux gradient, and take the absolute value of the dot product; The physical instability index is obtained by weighted summation of the absolute values ​​calculated for all beams.

[0014] Preferably, in step S160, the step of generating a predicted failure risk index by fusing the physical instability index and the anatomical-kinematic coupling index includes: The minimum-maximum normalization algorithm is used to map the physical instability index and the anatomical motion coupling index to a dimensionless interval, respectively, to obtain the normalized physical instability index and the normalized anatomical motion coupling index. The predicted failure risk index is obtained by linearly weighting and summing the normalized physical instability index and the normalized anatomical motion coupling index based on the physical factor weighting coefficient and the anatomical factor weighting coefficient.

[0015] Preferably, in step S160, the step of generating a predicted failure risk index by fusing the physical instability index and the anatomical-kinematic coupling index includes: The minimum-maximum normalization algorithm is used to map the physical instability index and the anatomical motion coupling index to a dimensionless interval, respectively, to obtain the normalized physical instability index and the normalized anatomical motion coupling index. The predicted failure risk index is obtained by linearly weighting and summing the normalized physical instability index and the normalized anatomical motion coupling index based on the physical factor weighting coefficient and the anatomical factor weighting coefficient.

[0016] Preferably, in step S160, the step of overlaying the confidence map as a heatmap onto the planned CT image and the initial dose prediction result includes: Using color coding technology, a mapping relationship from the confidence value to the color spectrum is established, and the confidence map is converted into a color confidence layer; Set the blending transparency parameter, and use alpha blending technology to display the color confidence layer, the planned CT image, and the initial dose prediction result in a semi-transparent overlay.

[0017] This invention provides an AI-based dose prediction-assisted analysis method for intensity-modulated radiotherapy (IMRT) of thoracic tumors. It offers the following advantages: 1. This invention addresses the challenge of quantifying dose delivery errors caused by the interaction between the dynamic target area and beam modulation in intensity-modulated radiotherapy (IMRT) of thoracic tumors by constructing the projection relationship of a three-dimensional deformation vector field onto the beam plane and combining it with a two-dimensional flux gradient to calculate the physical instability index. This method quantifies the flux cutting effect at the physical mechanism level, accurately locates high-risk areas for prediction failure, provides crucial physical features for AI dose prediction, and significantly improves the depth and accuracy of IMRT-assisted analysis for moving tumors.

[0018] 2. This invention addresses the problem of dose calculation deviation caused by dynamic changes in tissue density due to respiratory motion in thoracic tumor radiotherapy by calculating the anatomical motion coupling index between the three-dimensional deformation vector field and the three-dimensional electron density gradient field. This method captures the voxel characteristics moving along the direction of drastic density gradient changes, effectively quantifies the anatomical uncertainty caused by changes in water-equivalent depth, and compensates for the weakness of AI dose prediction models in perceiving dynamic anatomical structural changes, thereby improving the reliability of intensity-modulated radiotherapy (IMRT) auxiliary analysis methods in heterogeneous lung tissues.

[0019] 3. This invention generates a predicted failure risk index by fusing physical and anatomical features, and constructs a visualized confidence map overlaid on the planned CT image, establishing an intuitive AI dose prediction reliability assessment mechanism. This technical solution transforms the abstract motion coupling risk into a spatially distributed thermal layer, assisting physicists in quickly identifying high-risk areas that, although predicted, are severely affected by respiratory motion interference. This reduces clinical decision-making risk while ensuring the efficiency of AI applications, and greatly improves the quality control process for intensity-modulated radiotherapy (IMRT) analysis of thoracic tumors. Attached Figure Description

[0020] Figure 1 This is a flowchart of the AI-based dose prediction-assisted analysis method for intensity-modulated radiotherapy of thoracic tumors according to the present invention. Figure 2 This is the logic diagram for calculating the physical instability index of the present invention; Figure 3 This is a logic diagram for the multidimensional feature fusion and confidence visualization of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] This invention provides an AI-based dose prediction-assisted analysis method for intensity-modulated radiotherapy (IMRT) of thoracic tumors. The method is applied to computer equipment with data processing capabilities, such as a radiotherapy planning system (TPS) workstation, a medical image processing server, or a quality control workstation. The aforementioned computer equipment includes a processor, a memory, and a display unit. The memory stores a computer program, and the processor executes this computer program to implement the various steps of this method.

[0023] See attached document Figure 1 This AI-based dose prediction-assisted analysis method for intensity-modulated radiotherapy (IMRT) of thoracic tumors includes the following steps: S110: Obtain the four-dimensional CT image sequence, planned CT image, intensity-modulated radiotherapy planning file, and initial dose prediction results generated by the artificial intelligence model for the case to be analyzed; S120 parses the isocenter location information in the intensity-modulated radiotherapy (IMRT) planning file, establishes a treatment coordinate system with the isocenter as the origin, and uniformly converts the voxel coordinates of the four-dimensional CT image sequence and the planning CT image to the treatment coordinate system. S130 calculates the three-dimensional electron density gradient field based on the planned CT images and calculates the three-dimensional deformation vector field based on the four-dimensional CT image sequence, and performs consistency cleaning and regularization processing on the three-dimensional deformation vector field. S140: Analyze the beam parameters in the intensity-modulated radiotherapy (IMRT) planning file, generate a two-dimensional flux map of the corresponding beam viewpoint, calculate the two-dimensional flux gradient, construct the projection relationship of the three-dimensional deformation vector field on the beam plane based on the treatment coordinate system, and calculate the physical instability index in combination with the two-dimensional flux gradient. S150, based on the three-dimensional deformation vector field and the three-dimensional electron density gradient field, calculates the anatomical motion coupling index; S160 generates a predicted failure risk index by fusing the physical instability index and the anatomical motion coupling index, and determines the confidence map of the AI ​​dose prediction. The confidence map is then overlaid on the planned CT image and the initial dose prediction results in the form of a heatmap for auxiliary analysis.

[0024] See attached document Figure 1 In the specific implementation process, steps S110 to S120 mainly complete the acquisition of data and the unification of spatial reference, which is the basis for ensuring the geometric accuracy of subsequent multidimensional field coupling calculations.

[0025] In step S110, the processor obtains the complete data set of the case to be analyzed through the DICOM communication protocol or file reading interface. The complete data set includes at least the four-dimensional CT image sequence of the case to be analyzed, the planned CT image, the intensity-modulated radiotherapy plan file, and the initial dose prediction results generated by the artificial intelligence model.

[0026] Specifically, the four-dimensional CT image sequence contains anatomical images of the patient at different phases within a complete respiratory cycle. In this embodiment, the image at the end of inspiration is selected. and end-expiratory phase image As a key data frame characterizing the maximum respiratory motion amplitude, the planned CT image. The average density projection of a four-dimensional CT sequence is typically used as the anatomical reference for dose calculation. Intensity-modulated radiotherapy (IMRT) planning files contain physical parameters such as multi-leaf grating sequence, gantry angle, and source axis distance. The planned CT images are then processed by calling a pre-built artificial intelligence model (e.g., a deep neural network based on the U-Net architecture, which is supervised and trained using a large number of historical case anatomical images and clinically approved dose distribution maps). Input the AI ​​model and obtain the voxel-level three-dimensional dose matrix output by the AI ​​model as the initial dose prediction result. The parsing of DICOM data and the conversion of image formats are standard techniques in the field of medical image processing. Those skilled in the art can implement these techniques using open-source libraries (such as DCMTK or GDCM), and will not be elaborated upon here.

[0027] In step S120, spatial coordinate alignment is performed. By parsing the isocenter position information in the intensity-modulated radiotherapy (IMRT) plan file, the coordinate vector of the isocenter in the image coordinate system is extracted, denoted as... Subsequently, a treatment coordinate system with the isocenter as the origin is constructed. In this treatment coordinate system, the origin (0, 0, 0) corresponds to the physical treatment center. Due to the discrete nature of CT image data, the processor needs to handle the non-integer correspondence between voxel centers and physical coordinates during coordinate transformation. The processor employs a trilinear interpolation algorithm to ensure the numerical accuracy after coordinate transformation. By traversing each voxel in the four-dimensional CT image sequence and the planned CT image, the coordinate vector of that voxel in the image coordinate system is read and denoted as... And convert it into a coordinate vector in the treatment coordinate system. The coordinate transformation calculation formula is as follows: ; in, is the coordinate vector of the voxel in the image coordinate system; The coordinate vector of the isocenter in the image coordinate system; This represents the coordinate vector of the voxels in the treatment coordinate system after transformation. Through this transformation, the voxel coordinates of the four-dimensional CT image sequence and the planning CT image are uniformly transformed to the treatment coordinate system.

[0028] See attached document Figure 1In the specific implementation process, step S130 is mainly responsible for constructing a density field that represents the anatomical structure characteristics of the patient and a vector field that represents the motion characteristics of the organs, and for enhancing the artifact problems common in four-dimensional images to ensure the physical authenticity of the input data.

[0029] Calculating the three-dimensional electron density gradient field based on planned CT images. Specifically, this involves: pre-configuring a conversion protocol between CT values ​​and electron density, typically established using a bilinear fitting method. This protocol employs linear functions with different slopes in low-density regions (e.g., lungs) and high-density regions (e.g., bone) to more accurately describe the physical relationship between Hausfield units and relative electron density. A three-dimensional electron density volume is generated by converting the grayscale values ​​of each voxel in the planned CT image to electron density values. To quantify the drastic changes at anatomical interface points, a three-dimensional difference operator or the Sobel operator is applied to the three-dimensional electron density volume to calculate the spatial gradient vector of each voxel, thus obtaining the three-dimensional electron density gradient field, denoted as . The formula for calculating the three-dimensional electron density gradient field is as follows: ; in, The electron density value of a voxel; , , These are the partial derivatives of the electron density along the three axes of the treatment coordinate system. The high-value regions in this three-dimensional electron density gradient field correspond to the physical interfaces between lung tissue and tumor, heart, or chest wall.

[0030] The three-dimensional deformation vector field is calculated based on a four-dimensional CT image sequence. This is achieved using the end-expiratory phase image obtained in step S110. As the source image, the end-inspiratory phase image As the target image, a non-rigid registration operation is performed. This operation uses optical flow or B-spline free deformation algorithm to calculate the displacement vector of each voxel in the source image moving to its corresponding position in the target image, thereby generating the original three-dimensional deformation vector field, denoted as . .

[0031] To address potential non-physical singularities in the original 3D deformation vector field, a consistency cleaning and regularization process is performed. The divergence field of the original 3D deformation vector field, composed of the divergence values ​​of all voxels, is calculated and used to detect topological tearing or overlap caused by image artifacts. The formula for calculating the divergence value is as follows: ; in, , , They are respectively exist , , Component of direction; The divergence value of the voxels. Voxels whose absolute divergence value exceeds a preset physical threshold (e.g., a limit of 0.5 set based on the maximum compressibility of lung tissue) are identified as outliers, and the normal voxels in the neighborhood of the outlier are used to replace them, thus completing the consistency cleaning.

[0032] Based on consistent cleaning, regularization is performed. An anatomically guided filtering algorithm is employed to plan CT images. As a guide image, a smoothing constraint is applied to the cleaned 3D deformation vector field. The algorithm establishes a locally linear model, ensuring the output 3D deformation vector field remains smooth in flat regions of the anatomical structure, while preserving motion discontinuities in the edge regions. After this processing, the final 3D deformation vector field is output. This three-dimensional deformation vector field possesses both physical continuity and anatomical structural consistency.

[0033] See attached document Figure 1 and Figure 2 In the specific implementation process, step S140 mainly focuses on quantifying the dose delivery error caused by the interaction between beam intensity modulation and tumor movement in the patient during intensity-modulated radiotherapy. This step reduces the motion field in three-dimensional space to a two-dimensional plane from the beam's perspective through geometric projection transformation, and couples the analysis with the flux characteristics of this plane.

[0034] The beam parameters in the intensity-modulated radiotherapy (IMRT) planning file are parsed to generate a two-dimensional flux map of the corresponding beam angle and calculate the two-dimensional flux gradient. By traversing each beam in the IMRT planning file, its gantry angle, collimator angle, source axis distance, and control point sequence including the position of the multi-leaf grating blades are extracted. For each beam, a two-dimensional grid plane perpendicular to the beam's central axis is constructed, defining a two-dimensional coordinate system. ,in The axis corresponds to the direction of motion of the multi-leaf grating blades. The axis is perpendicular to the blade's direction of motion. By calculating the cumulative radiative flux intensity of each pixel on this two-dimensional grid plane based on the blade position data and corresponding subfield weights in the control point sequence, a two-dimensional flux map is generated, denoted as [image of the graph]. This cumulative calculation reflects the intensity distribution of the beam in a cross section perpendicular to the propagation direction throughout the entire treatment fraction.

[0035] Next, convolution is performed on the two-dimensional flux map to obtain the two-dimensional flux gradient. In this embodiment, a two-dimensional difference operator or the Sobel operator is used as the convolution kernel. Perform convolution operations to obtain the two-dimensional flux gradient, denoted as . The formula for calculating the two-dimensional flux gradient is as follows: ; in, and Flux intensity at shaft and The spatial rate of change along the axis, this gradient vector characterizes the direction in which flux intensity changes the most rapidly.

[0036] Based on the treatment coordinate system, the projection relationship of the three-dimensional deformation vector field onto the beam plane is constructed. This is because the three-dimensional deformation vector field generated in step S130... It is a three-dimensional treatment coordinate system defined with the isocenter as the origin. In this context, the flux gradient is defined in a two-dimensional coordinate system based on the beam's perspective. In this process, a spatial mapping is established between the two. This is done by adjusting the beam's gantry angle. and collimator angle Construct a 3D to 2D projection matrix This projection matrix is ​​used to rotate the motion vector in three-dimensional space to the beam viewpoint and remove the depth motion component along the beam axis to obtain the projected motion vector. This projection process only retains the planar motion components that cause flux misalignment. Projection matrix and projected motion vector The calculation formula is as follows: ; in, , , For three-dimensional deformation vector field Components in the treatment coordinate system; , Projected motion vector The components in the beam plane coordinate system. This projection calculation is performed traversing every voxel in the three-dimensional deformation vector field.

[0037] Finally, the physical instability index is calculated using the two-dimensional flux gradient. This index characterizes the degree of coupling between the anatomical motion direction and the beam flux gradient direction. It is based on the following physical principle: when the voxel's motion direction is perpendicular to the flux contour lines (i.e., parallel to the gradient direction), the voxel will traverse different dose intensity zones, leading to the maximum dose calculation error; conversely, if the motion direction is parallel to the flux contour lines, the error is minimal. This is achieved by calculating the projected motion vector. With two-dimensional flux gradient The dot product of the values ​​is calculated, and the absolute value of the dot product is taken. The weighted sum of the absolute values ​​calculated for all beams is then used to obtain the physical instability index. The calculation formula is as follows: ; in, The total number of beams; For beam index; For the first The weighting coefficient of each beam (determined based on the proportion of machine hops for that beam). This is for absolute value operations.

[0038] Calculated physical instability index It is a three-dimensional scalar field with the same dimension as the planned CT image space. The higher the value, the more severe the flux cutting effect that the region suffers during respiratory motion.

[0039] See attached document Figure 1 and Figure 3 In the specific implementation process, steps S150 to S160 mainly complete the deep coupling analysis of multidimensional field data and transform the analysis results into an intuitive confidence map to evaluate the predictive reliability of the artificial intelligence model in a specific anatomical region.

[0040] In step S150, the anatomical motion coupling index is calculated based on the three-dimensional deformation vector field and the three-dimensional electron density gradient field. This step aims to quantify the specific impact of tissue density changes caused by the patient's respiratory motion on the physical properties of the radiation penetration path. In radiotherapy physics, the deposition distribution of radiation energy mainly depends on the electron density integral (i.e., the water equivalent depth) along the radiation path. If the direction of human tissue movement is parallel to the direction of the electron density gradient (e.g., the edge of a lung tumor moves along a direction of drastic density change), it will cause a significant change in the water equivalent depth along the radiation path. This dynamic density change is the main anatomical factor causing deviations in the static dose calculation model; conversely, if the direction of movement is parallel to the iso-electron density surface, the effect on dose distribution is negligible.

[0041] By traversing each voxel within the computational region, the three-dimensional deformation vector field at that voxel's location is obtained. Vector values ​​and three-dimensional electron density gradient field The gradient value is calculated. The dot product of the vector value and the gradient value is calculated, and the absolute value of the dot product is taken. The result is used as the anatomical motion coupling index of the voxel position, denoted as . The formula for calculating the anatomical-motor coupling index is as follows: ; in, , , For three-dimensional deformation vector field The component at that point; , , Three-dimensional electron density gradient field The component at that point; This represents absolute value operations. The larger the value, the more drastic the change in medium density the voxel location experienced during respiratory movements, which constitutes a source of anatomical uncertainty that is difficult for artificial intelligence models to capture.

[0042] In step S160, a predicted failure risk index is generated by fusing the physical instability index and the anatomical motion coupling index. Due to the physical instability index... Coupling index with anatomical movement Since these two exponential fields have different physical dimensions and numerical ranges, they are first normalized separately. By employing a min-max normalization algorithm, the two exponents are mapped to the dimensionless interval [0, 1] to eliminate the dimensional differences in the original data, resulting in the normalized physical instability exponents. and normalized anatomical-motor coupling index .

[0043] Subsequently, a linear weighted fusion algorithm was used to calculate the predicted failure risk index, denoted as . The calculation formula is as follows: ; in, and These are the weighting coefficients for the physical factor and the weighting coefficients for the anatomical factor, respectively, and satisfy the following conditions: These two coefficients are configured based on the specific characteristics of the treatment site. For example, in stereotactic radiotherapy of the lungs, due to the significant influence of density heterogeneity, the system will automatically adjust them higher. The value of . This comprehensively reflects the potential error risk in dose prediction caused by physical modulation effects and changes in anatomical structure.

[0044] Determine the confidence plot for the Al dose prediction. Confidence plot Defined as the inverse mapping of predicted failure risk, it is used to intuitively characterize the credibility of the output results of artificial intelligence models. Specifically, the process of generating a confidence map is a voxel-level numerical transformation of the three-dimensional predicted failure risk index field. By using the general sigmoid nonlinear mapping function, each voxel in the predicted failure risk index field is traversed, and the predicted failure risk index at that voxel position is converted into a confidence value in the interval (0, 1). The set of confidence values ​​calculated for all voxel positions constitutes the confidence map.

[0045] To enhance the warning effect of high-risk areas, the above conversion is based on the following calculation formula: ; in, voxel coordinates in the confidence plot Confidence value at; The slope parameter of the Sigmoid function controls the sensitivity of the confidence level as the risk increases. The risk threshold parameter defines the inflection point at which risk becomes acceptable. The risk level can be determined by analyzing the statistical distribution of prediction errors in historical retrospective cohorts, typically set as the risk value corresponding to the 95th quantile of the error distribution. This formula maps high-risk areas to low-confidence values ​​(values ​​approaching 0) and low-risk areas to high-confidence values ​​(values ​​approaching 1).

[0046] Finally, the confidence map is overlaid on the planned CT image and the initial dose prediction results as a heatmap. Using color coding technology, a mapping relationship is established from confidence values ​​to color spectra (e.g., red-yellow-green), converting the confidence map into a color confidence layer, where red corresponds to low confidence and green to high confidence. Using alpha blending technology, by setting a blending transparency parameter (e.g., 0.5), the generated color confidence layer is semi-transparently overlaid on the original planned CT image layer or initial dose prediction result layer. This display method allows users to intuitively identify high-risk areas (i.e., the red areas in the image) where the AI ​​model generated a dose distribution, but the prediction results are highly unstable due to complex respiratory motion and beam modulation coupling effects. This helps physicists focus on reviewing the dose compliance of these areas.

[0047] To better understand the technical solution of the present invention, the above embodiments will be described in detail below with reference to a specific case scenario of stereotactic radiotherapy for non-small cell lung cancer in the lower lobe of the right lung.

[0048] Scenario: The patient is a 65-year-old male with a solid tumor approximately 3 cm in diameter in the lower lobe of his right lung, accompanied by significant respiratory movements (diaphragmatic movement causing considerable displacement of the tumor in the head-to-foot direction). The physicist has developed a dynamic intensity-modulated radiotherapy (IMRT) plan comprising 5 beams. The method of this invention is run on a radiotherapy planning system workstation.

[0049] Step 1: Data Loading and Coordinate Unification The TPS workstation loads the patient's four-dimensional CT image sequence (containing 10 respiratory phases) and selects the end-inspiratory phase. and end-expiratory phase Simultaneously load planned CT images. and the initial dose distribution predicted by the AI ​​model The system analyzes the intensity-modulated radiotherapy (IMRT) plan file to determine the isocenter. Located at the geometric center of the tumor. By traversing all voxels, using the formula... The image coordinates are converted into a treatment coordinate system with the tumor center as the origin to ensure that the spatial reference for subsequent calculations is consistent.

[0050] Step 2: Construction and Cleaning of Field Data Electron density gradient calculation: At the tumor margin, due to the high density of tumor tissue (approximately 1.0 g / cm³), 3 ) and low-density alveolar tissue (approximately 0.25 g / cm³) 3 Adjacent to each other, the system uses the formula The calculations showed that there was a large gradient value at the tumor margin. The direction is pointing towards the inside of the tumor.

[0051] Deformation vector calculation and cleaning: System calculation from arrive The displacement generates the original three-dimensional deformation vector field. Due to the large amplitude of motion in the lower lobe of the right lung, non-physical folds (divergence anomalies) appeared near the diaphragm in the original field. The system calculates the divergence. The system detected a divergence value exceeding 0.5 in a certain region, identifying it as artifact noise. The system replaced this value with the neighborhood mean, and after anatomical-guided filtering and regularization, outputted a smooth, physiologically consistent three-dimensional deformation vector field. The tumor was found to have moved approximately 1.5 cm towards the base (negative Z-axis direction).

[0052] Step 3: Physical Instability Analysis (Beam Perspective) At one of the rack angles For example, a beam that shines vertically downwards: Flux gradient calculation: To protect the normal lung tissue beneath the tumor, the MLC blades create a steep intensity drop at the lower edge of the tumor. The system generates a two-dimensional flux map. The region changes drastically, resulting in a two-dimensional flux gradient. The mold is very long.

[0053] Projection and Coupling: The system is based on Constructing the projection matrix Since the tumor's main direction of motion (Z-axis) is parallel to the longitudinal axis (v-axis) of the beam plane, the three-dimensional deformation vector field... Projected motion vector on the beam plane It's very big.

[0054] Exponent calculation: Projected motion vector at the lower edge of the tumor Direction and two-dimensional flux gradient The directions are almost parallel (i.e., the direction of motion is perpendicular to the isodose lines, leading to severe "dose ambiguity" or "interaction effects"). The system is based on the formula... The physical instability index of the region was calculated. Extremely high.

[0055] Step 4: Anatomical-motor coupling analysis In the tumor periphery region, the tumor moves with respiration (three-dimensional deformation vector field). It passes through an area that was originally lung tissue. At this point, the direction of motion is aligned with the three-dimensional electron density gradient field. Highly parallel. The system is based on the formula. Calculations revealed that the anatomical-motor coupling index of this region... This is also very high. This means that the equivalent water depth through which the rays pass fluctuates dramatically during the breathing cycle, and the dose predicted by static Al may not reflect the actual dose deposition.

[0056] Step 5: Risk Assessment and Visualization Normalization and Fusion: The system normalizes the above two indices to and Given that SBRT is sensitive to density changes, the system is configured with anatomical weights. Physical weight Using formulas A high-risk index of nearly 0.9 was generated at the tumor margin (especially the lower margin). .

[0057] Confidence level generation: The system sets a risk threshold. slope Substitute into the formula High-risk areas (0.9) results in a maximal denominator, which in turn increases the confidence level. Close to 0 (low confidence level).

[0058] Final display: The physicist saw on the screen the planned CT images in grayscale. and color AI predictive dose Above this, a red heatmap ring was superimposed, which precisely covered the lower edge of the tumor.

[0059] Analysis conclusion: This intuitively suggests to physicists that although the AI ​​model predicts the dose to be within the target range, the actual dose at that location may be seriously deviated (e.g., insufficient or excessive) due to the coupling of intense respiratory movements and drastic density / flux changes.

[0060] Subsequent actions: Based on this red high-risk warning, the physicist decided to adopt an ITV (inner target volume) expansion strategy or re-optimize the MLC motion trajectory to smooth the flux gradient, thereby reducing treatment risks.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based dose prediction-assisted analysis method for intensity-modulated radiotherapy (IMRT) of thoracic tumors, characterized in that, Includes the following steps: S110: Obtain the four-dimensional CT image sequence, planned CT image, intensity-modulated radiotherapy planning file, and initial dose prediction results generated by the artificial intelligence model for the case to be analyzed; S120, parse the isocenter position information in the intensity-modulated radiotherapy planning file, establish a treatment coordinate system with the isocenter as the origin, and uniformly convert the voxel coordinates of the four-dimensional CT image sequence and the planning CT image to the treatment coordinate system; S130, calculate the three-dimensional electron density gradient field based on the planned CT image, and calculate the three-dimensional deformation vector field based on the four-dimensional CT image sequence, and perform consistency cleaning and regularization processing on the three-dimensional deformation vector field; S140, parse the beam parameters in the intensity-modulated radiotherapy planning file, generate a two-dimensional flux map corresponding to the beam viewpoint, and calculate the two-dimensional flux gradient; construct the projection relationship of the three-dimensional deformation vector field on the beam plane according to the treatment coordinate system, and calculate the physical instability index in combination with the two-dimensional flux gradient; S150, Calculate the anatomical motion coupling index based on the three-dimensional deformation vector field and the three-dimensional electron density gradient field; S160, by fusing the physical instability index and the anatomical motion coupling index, a predicted failure risk index is generated, and a confidence map of the AI ​​dose prediction is determined. The confidence map is then overlaid on the planned CT image and the initial dose prediction result in the form of a heatmap.

2. The method for auxiliary analysis of intensity-modulated radiotherapy for thoracic tumors based on AI dose prediction according to claim 1, characterized in that, In step S130, the step of calculating the three-dimensional electron density gradient field based on the planned CT image includes: By using a pre-configured CT value to electron density conversion protocol, the gray value of each voxel in the planned CT image is converted into an electron density value to generate a three-dimensional electron density volume. The three-dimensional electron density gradient field is obtained by applying a three-dimensional difference operator or a Sobel operator to the three-dimensional electron density volume to calculate the spatial gradient vector of each voxel point.

3. The method for auxiliary analysis of intensity-modulated radiotherapy for thoracic tumors based on AI dose prediction according to claim 1, characterized in that, In step S130, the step of calculating the three-dimensional deformation vector field based on the four-dimensional CT image sequence includes: The end-expiratory phase image in the four-dimensional CT image sequence is selected as the source image, and the end-inspiratory phase image is selected as the target image. A non-rigid registration operation is performed on the source image and the target image to calculate the displacement vector of each voxel in the source image moving to the corresponding position in the target image, thereby generating the original three-dimensional deformation vector field.

4. The AI-based dose prediction-assisted analysis method for intensity-modulated radiotherapy of thoracic tumors according to claim 3, characterized in that, In step S130, the step of performing uniform cleaning and regularization processing on the three-dimensional deformation vector field includes: Calculate the divergence field of the original three-dimensional deformation vector field; Voxels whose absolute divergence values ​​exceed a preset physical threshold are identified as outliers, and the outliers are replaced by the vector mean of normal voxels in the neighborhood of the outlier. An anatomical-guided filtering algorithm is used, with the planned CT image as the guiding image, to smooth and constrain the replaced three-dimensional deformation vector field.

5. The method for auxiliary analysis of intensity-modulated radiotherapy for thoracic tumors based on AI dose prediction according to claim 1, characterized in that, In step S140, the steps of parsing the beam parameters in the intensity-modulated radiotherapy (IMRT) planning file, generating a two-dimensional flux map of the corresponding beam angle, and calculating the two-dimensional flux gradient include: Traverse each beam in the intensity-modulated radiotherapy planning file and construct a two-dimensional grid plane perpendicular to the beam's central axis; Based on the control point sequence of the multi-leaf grating blade positions and the corresponding subfield weights, the cumulative radiative flux intensity of each pixel on the two-dimensional grid plane is calculated to generate the two-dimensional flux map. Perform convolution calculations on the two-dimensional flux graph to obtain the two-dimensional flux gradient.

6. The method for auxiliary analysis of intensity-modulated radiotherapy for thoracic tumors based on AI dose prediction according to claim 1, characterized in that, In step S140, the step of constructing the projection relationship of the three-dimensional deformation vector field onto the beam plane based on the treatment coordinate system, and calculating the physical instability index in conjunction with the two-dimensional flux gradient, includes: Based on the frame angle and collimator angle of the beam, a three-dimensional to two-dimensional projection matrix is ​​constructed; The three-dimensional deformation vector field is rotated to the beam viewpoint using the projection matrix, and the depth motion component along the beam axis is removed to obtain the projected motion vector; Calculate the dot product of the projected motion vector and the two-dimensional flux gradient, and take the absolute value of the dot product; The physical instability index is obtained by weighted summation of the absolute values ​​calculated for all beams.

7. The method for auxiliary analysis of intensity-modulated radiotherapy for thoracic tumors based on AI dose prediction according to claim 1, characterized in that, In step S150, the step of calculating the anatomical motion coupling index based on the three-dimensional deformation vector field and the three-dimensional electron density gradient field includes: Traverse each voxel within the computational region to obtain the vector value of the three-dimensional deformation vector field and the gradient value of the three-dimensional electron density gradient field at the voxel location; Calculate the dot product of the vector value and the gradient value, and take the absolute value of the dot product. Use the calculation result as the anatomical motion coupling index of the voxel position.

8. The method for auxiliary analysis of intensity-modulated radiotherapy for thoracic tumors based on AI dose prediction according to claim 1, characterized in that, In step S160, the step of generating a predicted failure risk index by fusing the physical instability index and the anatomical motion coupling index includes: The minimum-maximum normalization algorithm is used to map the physical instability index and the anatomical motion coupling index to a dimensionless interval, respectively, to obtain the normalized physical instability index and the normalized anatomical motion coupling index. The predicted failure risk index is obtained by linearly weighting and summing the normalized physical instability index and the normalized anatomical motion coupling index based on the physical factor weighting coefficient and the anatomical factor weighting coefficient.

9. The method for auxiliary analysis of intensity-modulated radiotherapy for thoracic tumors based on AI dose prediction according to claim 1, characterized in that, In step S160, the step of determining the confidence map of the AI ​​dose prediction includes: The predicted failure risk index is converted into a confidence value using the Sigmoid nonlinear mapping function to generate the confidence map; wherein, the parameters of the Sigmoid nonlinear mapping function include a slope parameter that controls the sensitivity and a risk threshold parameter that defines the risk inflection point.

10. The method for auxiliary analysis of intensity-modulated radiotherapy for thoracic tumors based on AI dose prediction according to claim 1, characterized in that, In step S160, the step of overlaying the confidence map as a heatmap onto the planned CT image and the initial dose prediction result includes: Using color coding technology, a mapping relationship from the confidence value to the color spectrum is established, and the confidence map is converted into a color confidence layer; Set the blending transparency parameter, and use alpha blending technology to display the color confidence layer, the planned CT image, and the initial dose prediction result in a semi-transparent overlay.

Images