An interactive radiotherapy planning system optimization method

Abstract

The application relates to the technical field of radiotherapy optimization, and discloses an interactive radiotherapy planning system optimization method. The method comprises the following steps: acquiring medical image data and real-time body surface motion monitoring data of a patient, constructing a three-dimensional dose flux distribution map according to the medical image data, and generating an anatomical displacement feature set according to the real-time body surface motion monitoring data; inputting the three-dimensional dose flux distribution map and the anatomical displacement feature set into a pre-constructed multi-modal dose prediction model to obtain a first prediction result of dose distribution deviation; extracting radiation biological feature parameters from the three-dimensional dose flux distribution map, matching the radiation biological feature parameters with a pre-set radiation damage threshold model to obtain a second prediction result of dose safety evaluation; and generating dynamic optimization weights based on the first prediction result and the second prediction result, and adjusting treatment planning parameters according to the dynamic optimization weights to realize radiotherapy optimization.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy optimization technology, specifically to an interactive radiotherapy planning system optimization method. Background Technology

[0002] In the field of radiotherapy, the accuracy of treatment planning directly affects treatment efficacy and patient safety. However, as radiotherapy technology advances towards higher precision and higher doses, traditional radiotherapy planning systems are increasingly revealing numerous shortcomings. Currently, most radiotherapy planning systems rely primarily on static medical imaging data to construct dose distribution models when developing treatment plans, failing to adequately consider the impact of the patient's physiological movements during treatment. During treatment, patients may experience surface movements such as breathing, heartbeat, and changes in body position. These movements can cause shifts in the position of internal anatomical structures, resulting in deviations between the actual dose distribution and the pre-planned dose distribution. This can lead to insufficient dose to the tumor target area, affecting treatment efficacy, and may also result in excessive radiation to surrounding normal tissues, increasing the risk of complications. To address the issue of anatomical displacement, some radiotherapy planning systems have attempted to incorporate surface motion monitoring data. However, existing fusion methods often remain at the level of simple data overlay, failing to construct effective multimodal data association models and thus unable to accurately predict dose distribution deviations caused by anatomical displacement. Furthermore, in dose safety assessment, traditional methods typically rely solely on a single dose value, lacking consideration of radiation biology characteristics and failing to comprehensively reflect the differences in radiation sensitivity among different tissues and the potential risks of radiation damage. For example, the effects of the same dose of radiation on tumor tissue and normal organs differ significantly. Using only dose thresholds cannot accurately assess the safety of treatment plans, potentially leading to treatment plans that prioritize tumor control while neglecting the protection of normal tissues. Current radiotherapy planning systems are mostly optimized statically, meaning they are based on pre-treatment data to create a fixed treatment plan. During treatment, they cannot dynamically adjust optimization weights and treatment parameters according to real-time monitoring of anatomical displacement changes and dose distribution deviations. This static optimization model struggles to adapt to dynamic changes during treatment, further reducing the accuracy and safety of the treatment plan. With the continuous development of radiotherapy technology, clinicians have placed higher demands on the accuracy, dynamic adaptability, and comprehensive safety assessment of treatment planning systems. There is an urgent need for a radiotherapy planning system optimization method that can effectively integrate multimodal data and achieve dynamic optimization and accurate safety assessment to address the aforementioned problems in existing technologies. Summary of the Invention

[0003] The purpose of this invention is to provide an interactive radiotherapy planning system optimization method to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides an interactive radiotherapy planning system optimization method, the method comprising: Acquire the patient's medical imaging data and real-time body surface motion monitoring data, construct a three-dimensional dose-flux distribution map based on the medical imaging data, and generate an anatomical displacement feature set based on the real-time body surface motion monitoring data; The three-dimensional dose flux distribution map and the anatomical displacement feature set are input into a pre-constructed multimodal dose prediction model for processing to obtain the first prediction result of the dose distribution deviation; Radiation biological characteristic parameters are extracted from the three-dimensional dose flux distribution map and matched with a preset radiation damage threshold model to obtain a second prediction result for dose safety assessment. Dynamic optimization weights are generated based on the first prediction result and the second prediction result, and the treatment plan parameters are adjusted according to the dynamic optimization weights to achieve radiotherapy optimization.

[0005] Preferably, the step of constructing a three-dimensional dose-flux distribution map based on the medical imaging data and generating an anatomical displacement feature set based on the real-time body surface motion monitoring data includes: The medical image data is used to perform dose flux simulation calculations to generate an initial three-dimensional dose flux distribution map, and key dose gradient features are extracted from the initial three-dimensional dose flux distribution map; The spatiotemporal correlation between the real-time body surface motion monitoring data and the key dose gradient features is quantified to generate an anatomical displacement vector field; The key dose gradient features and the anatomical displacement vector field are standardized and integrated in a time series format to generate the anatomical displacement feature set.

[0006] Preferably, quantifying the spatiotemporal correlation between the real-time body surface motion monitoring data and the key dose gradient features includes: Phase alignment analysis was performed on the synchronously acquired real-time body surface motion monitoring data and the key dose gradient features to calculate the displacement-dose coupling coefficient; Based on the distribution density of the key dose gradient features, the displacement-dose coupling coefficient is spatially weighted. The weighted displacement-dose coupling coefficients are mapped to the organ motion trajectory coordinate system to generate a three-dimensional anatomical displacement vector field.

[0007] Preferably, the multimodal dose prediction model includes an image encoder, a motion feature encoder, and a multimodal fusion decoder; The step of inputting the three-dimensional dose flux distribution map and the anatomical displacement feature set into a pre-constructed multimodal dose prediction model for processing includes: The dose flux feature vector of the three-dimensional dose flux distribution map is extracted by the image encoder; The spatiotemporal motion feature vector of the anatomical displacement feature set is parsed by the motion feature encoder. The dose flux feature vector and the spatiotemporal motion feature vector are input into the multimodal fusion decoder based on the cross-modal attention mechanism, and the first prediction result of the dose distribution deviation is output.

[0008] Preferably, the step of extracting radiation biological characteristic parameters based on the three-dimensional dose flux distribution map includes: Identify the dose hotspot distribution of high-risk organs in the three-dimensional dose flux distribution map; Calculate the rate of change of the slope of the cumulative dose curve for the dose hotspot distribution of the high-risk organs; Extract isodose curvature features and dose drop gradient features of dose-sensitive organs; The rate of change of the slope of the dose accumulation curve, the curvature characteristics of the isodose line, and the dose drop gradient characteristics are combined into a set of radiation biological characteristic parameters.

[0009] Preferably, the matching with a preset radiation damage threshold model includes: The radiation biology feature parameter set is matched with a preset organ tolerance dose database for feature similarity. The probability value of radiation damage to each high-risk organ is calculated based on the matching similarity. The second prediction result is generated based on the degree of deviation between the radiation damage probability value and the preset safety threshold.

[0010] Preferably, the step of generating dynamic optimization weights based on the first prediction result and the second prediction result includes: The dose deviation confidence level in the first prediction result is extracted as the first weighting factor; The organ damage risk level in the second prediction result is used as the second weighting factor; The first weighting factor and the second weighting factor are subjected to nonlinear normalization to generate dynamic optimized weights for organ protection priority.

[0011] Preferably, adjusting the treatment plan parameters according to the dynamically optimized weights includes: When the anatomical displacement feature set shows that the organ displacement exceeds the threshold, a multi-level dose compensation strategy based on the dynamic optimization weight is activated. The beam angle weight distribution is reconstructed based on the multi-level dose compensation strategy; Adjust the blade motion trajectory parameters and dose rate control curve of the multi-leaf collimator.

[0012] Preferably, activating the multi-level dose compensation strategy based on the dynamically optimized weights includes: Based on the organ displacement amplitude, compensation levels are divided, and a multi-level dose gradient compensation layer is constructed. Obtain the deviation between real-time dose monitoring data and the planned dose; The compensation intensity parameters of the multi-level dose gradient compensation layer are dynamically adjusted based on the deviation.

[0013] Preferably, after adjusting the blade motion trajectory parameters and dose rate control curve of the multi-leaf collimator, the method further includes: Generate a treatment execution risk feature list, which includes organ displacement risk markers and dose hot and cold spot distribution markers; The training dataset of the multimodal dose prediction model is updated according to the risk feature list; The fusion decision parameters of the multimodal fusion decoder are iteratively optimized based on the updated training dataset.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This interactive radiotherapy planning system optimization method offers an effective solution for improving the accuracy and safety of radiotherapy planning through multi-dimensional technological innovation. Firstly, at the data fusion level, the method simultaneously acquires the patient's medical imaging data and real-time body surface motion monitoring data. Based on the medical imaging data, a three-dimensional dose-flux distribution map is constructed, and combined with the real-time body surface motion monitoring data, an anatomical displacement feature set is generated, achieving effective integration of static anatomical information and dynamic motion information. This multi-source data fusion approach breaks through the limitations of the separation of static and dynamic data in traditional systems, and can more comprehensively reflect the actual state of the patient's anatomical structure and dose distribution during treatment, laying the foundation for subsequent dose prediction and optimization. In terms of dose distribution deviation prediction, this method inputs a three-dimensional dose flux distribution map and an anatomical displacement feature set into a pre-constructed multimodal dose prediction model for processing, obtaining the first prediction result of dose distribution deviation. The multimodal dose prediction model can fully explore the correlation between different types of data. By establishing a mapping relationship between anatomical displacement and dose distribution deviation, it accurately predicts dose distribution changes caused by anatomical structure displacement, significantly improving the accuracy of dose distribution deviation prediction compared to the traditional simple data overlay method. With this prediction result, medical staff can anticipate potential dose deviations during treatment, providing direction for adjusting treatment parameters and avoiding insufficient dose to the tumor target area or excessive irradiation of normal tissues due to dose deviation. In the dose safety assessment stage, this method extracts radiation biological characteristic parameters based on a three-dimensional dose flux distribution map and matches them with a pre-defined radiation damage threshold model to obtain a second predictive result for dose safety assessment. Radiation biological characteristic parameters reflect the sensitivity of different tissues to radiation and the potential risk of radiation damage. Combined with the radiation damage threshold model, this allows for a multi-dimensional assessment of the safety of treatment protocols, rather than relying solely on traditional dose values. This assessment method fully considers the differences in tissue radiation sensitivity, enabling a more comprehensive identification of potential safety hazards in the treatment plan. For example, while ensuring sufficient dose to the tumor target area, it effectively assesses the radiation damage risk to surrounding normal tissues, achieving a balance between tumor treatment and protection of normal tissues. Regarding dynamic optimization of treatment plans, this method generates dynamic optimization weights based on the first and second prediction results, and adjusts the treatment plan parameters according to these weights. The generation of dynamic optimization weights allows for flexible adjustment of the priority of each optimization objective based on real-time monitored dose distribution deviations and safety assessment results. For example, when a potential dose insufficiency in the tumor target area is predicted, the weight of tumor target area dose coverage can be appropriately increased; when a high radiation risk to normal tissues is detected, the weight of normal tissue protection can be increased. This dynamic adjustment mechanism overcomes the limitations of traditional static optimization models, enabling the treatment plan to adapt in real-time to changes in anatomical displacement and dose distribution deviations during treatment, ensuring that the treatment plan maintains high accuracy and safety throughout the entire treatment process. This method, through the organic integration of various stages, forms a complete interactive optimization system, achieving closed-loop optimization from data acquisition, deviation prediction, safety assessment to parameter adjustment. This interactive optimization mode allows medical staff to more intuitively grasp the dynamic changes in the treatment plan and participate in the optimization process according to actual needs, improving the flexibility and personalization of treatment plan formulation. It can better meet the treatment needs of different patients, and is especially suitable for patients with significant anatomical structure movement or high radiation sensitivity, further expanding the applicability and clinical application value of radiotherapy planning systems. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the working principle of the interactive radiotherapy planning system optimization method described in this invention. Figure 2 Flowchart for constructing a three-dimensional dose flux distribution map and generating an anatomical displacement feature set; Figure 3 Flowchart for inputting three-dimensional dose flux distribution map and anatomical displacement feature set into multimodal dose prediction model processing; Figure 4 Flowchart for matching radiation biological characteristic parameters with a preset radiation damage threshold model. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0017] Please see Figure 1 This invention provides a method for optimizing an interactive radiotherapy planning system, the method comprising: Acquire the patient's medical imaging data and real-time body surface motion monitoring data, construct a three-dimensional dose-flux distribution map based on the medical imaging data, and generate an anatomical displacement feature set based on the real-time body surface motion monitoring data; The three-dimensional dose flux distribution map and the anatomical displacement feature set are input into a pre-constructed multimodal dose prediction model for processing to obtain the first prediction result of the dose distribution deviation; Radiation biological characteristic parameters are extracted from the three-dimensional dose flux distribution map and matched with a preset radiation damage threshold model to obtain a second prediction result for dose safety assessment. Dynamic optimization weights are generated based on the first prediction result and the second prediction result, and the treatment plan parameters are adjusted according to the dynamic optimization weights to achieve radiotherapy optimization.

[0018] Example 1: See Figure 2 Medical imaging data contains anatomical information about patients acquired through equipment such as computed tomography (CT), magnetic resonance imaging (MRI), or positron emission tomography (PET). This data exists in the form of digital images, typically containing spatial resolution and grayscale information. Before constructing a three-dimensional dose-flux map, necessary preprocessing operations are performed on the raw medical imaging data. Preprocessing includes image registration, aligning image data acquired at different time points or modalities to a common reference coordinate system. Subsequently, image segmentation is performed, using thresholding, region growing, or atlas-based segmentation algorithms to distinguish and label different anatomical structures, especially clinical target areas requiring radiation exposure, organs at risk of overexposure such as the spinal cord, optic nerve, and kidneys, as well as surrounding normal tissues. The segmentation results define the spatial extent of different tissue structures.

[0019] After segmentation, the preprocessed medical image data, along with the planned radiation physics parameters, are input into the dose calculation engine built into the radiotherapy planning system. These radiation physics parameters include accelerator beam energy selection, field size setting, beam angle configuration, wedge filter selection, and multi-leaf collimator blade position sequence files. Based on the input physics parameters and the patient's specific tissue density distribution information, the dose calculation engine uses a computational model to simulate the transport process and energy deposition of radiation particles within human tissue. Commonly used computational models include detailed particle tracking simulations based on the Monte Carlo method, or approximate algorithms based on the pencil beam convolution superposition principle. After execution, the engine outputs dose distribution data on a three-dimensional spatial grid. This data records in detail the amount of radiation energy deposited in each spatial voxel, as calculated in the simulation. The calculated dose values ​​are mapped back to the spatial coordinate system of the original image to form an initial three-dimensional dose flux distribution map. This distribution map is a digital model characterizing the three-dimensional spatial distribution of the planned radiation dose within the patient's body.

[0020] Feature extraction is performed on the generated initial 3D dose flux distribution map to identify key regions with significant dose changes. The rate of change of the dose distribution map along the three coordinate axes at each spatial location is calculated using a spatial gradient operator. The gradient magnitude is calculated, which is a scalar value obtained by combining the rates of change in the three directions. Regions with high gradient magnitudes represent locations where the dose changes drastically in space. The analysis focuses on spatial locations where the gradient magnitude exceeds a preset threshold. These locations typically correspond to two types of key regions: first, the boundary between high-dose and low-dose regions, i.e., the edge of the dose drop zone; and second, the boundary of local maximum points or small-scale high-dose clusters in the dose distribution, i.e., the edge of dose hotspots. The spatial coordinate set of these high-gradient-magnitude regions and their corresponding dose gradient direction vector information are extracted; this information together constitutes the key dose gradient features. The key dose gradient features quantify the locations and directions in the planned dose distribution that are most sensitive to spatial changes.

[0021] Meanwhile, real-time motion monitoring data of the patient's body surface is continuously collected through optical surface monitoring systems installed in the treatment room or marker tracking devices attached to the patient's skin. The optical surface monitoring system captures the contours of the patient's body surface using multiple cameras, reconstructs the three-dimensional shape of the surface, and tracks its changes over time. The marker tracking device accurately records the three-dimensional coordinate changes of reflective or electromagnetic markers attached to the patient's skin. These devices continuously output a high-temporal-resolution data stream, reflecting the movement trajectory of the patient's body surface during free breathing or other physiological activities, forming a real-time body surface motion monitoring dataset.

[0022] The core step is to establish a quantitative correlation between real-time surface motion monitoring data and the previously extracted key dose gradient features, revealing how surface motion affects changes in internal dose distribution. The synchronicity of the two in both temporal and spatial dimensions is analyzed. Temporally, time-series data of surface motion monitoring points and key dose gradient features are aligned using timestamps. Spatially, a spatial correspondence between surface location points and internal anatomical structure points is established. This relationship can be established through the association information between surface markers and internal anatomical reference points during patient positioning, or through a surface-to-internal displacement field mapping model obtained based on deformation image registration. Based on the spatiotemporally aligned data, a statistical correlation index, the displacement-dose coupling coefficient, is calculated between the change in motion displacement at each surface monitoring point or surface region and the change in dose gradient at its associated internal key dose gradient feature point. This coefficient quantifies the extent to which a specific surface motion pattern drives changes in dose distribution at a specific internal location. The calculation process is performed for different time points and spatial locations.

[0023] Because critical dose gradient features are not uniformly distributed spatially, certain regions, such as high-gradient areas near organs at risk, have a greater impact on the accuracy of the overall dose distribution. Therefore, the calculated original displacement-dose coupling coefficients are spatially weighted. The weight values ​​are dynamically allocated based on the spatial distribution density function of the critical dose gradient features. Regions with high density of critical dose gradient feature points, such as areas where multiple dose hotspots or drop zones converge, or areas near highly radiation-sensitive organs, are assigned higher weight coefficients. The spatially weighted displacement-dose coupling coefficients are transformed into a unified organ motion trajectory coordinate system based on a mapping function of the organ motion trajectory model. In this coordinate system, spatial interpolation techniques such as radial basis function interpolation or Kriging interpolation are used to fit the discrete, weighted displacement-dose coupling coefficient sample points, generating a continuous three-dimensional vector field covering the entire anatomical region of interest, i.e., the anatomical displacement vector field. Each vector in this vector field represents the predicted three-dimensional displacement direction and amplitude of the anatomical structure caused by body surface movement at that spatial location. Anatomical displacement vector field transforms surface motion information into predictions of displacement of internal anatomical structures.

[0024] To ensure the extracted critical dose gradient features and the generated anatomical displacement vector field are effectively processed and comparable by subsequent models, data standardization is necessary. Standardization operations include scale normalization, which maps the amplitudes of the critical dose gradient features and the displacement values ​​of the anatomical displacement vector field to a standard interval of zero to one, eliminating the influence of dimensions; spatial coordinate normalization, which transforms all spatial coordinates to the same standardized three-dimensional coordinate system, typically the coordinate system of the reference image; and temporal alignment, which ensures that the critical dose gradient feature data and the anatomical displacement vector field data have the same sampling frequency and time point labels in the time series. For cases of inconsistent time points, linear interpolation methods can be used for compensation.

[0025] The standardized key dose gradient feature data and anatomical displacement vector field data are integrated and encapsulated according to a strict time-series structure. Each time point corresponds to a data packet, which contains two core parts: the first part is the spatial distribution information of the key dose gradient features at that time point, i.e., the position coordinates of all high gradient amplitude points and their normalized gradient direction vectors; the second part is the anatomical displacement vector field data at that time point, i.e., the three-dimensional displacement vector information of the predicted internal anatomical structures in the standard coordinate system. All data packets at all time points are arranged in chronological order to form the final anatomical displacement feature set. This dataset completely encapsulates the time-evolving dose distribution sensitive region information and the predicted anatomical structure displacement information, providing structured input for subsequent multimodal dose prediction models.

[0026] Example 2: See Figure 3 Real-time body surface motion monitoring data and key dose gradient feature data are acquired in parallel as two independent time-series data sources. The former originates from an optical surface scanning system or a body surface marker tracking device, recording the three-dimensional positional changes of specific points or regions on the patient's body surface at a fixed sampling frequency. The latter is dynamically extracted from a time-updated three-dimensional dose flux distribution map, reflecting information on spatially sensitive areas in the planned dose distribution. These two data streams need to be precisely synchronized in time. A timestamp alignment method based on a common time reference is used. If periodic physiological signals such as respiratory signals exist, respiratory phase gating technology is used for phase alignment analysis, segmenting the two data streams into the same phase window of the respiratory cycle. Based on strict time alignment, for each spatial location or associated body surface-internal spatial corresponding point, a statistical correlation measure is calculated between the change in displacement vector of the body surface location point or region and the change in dose gradient amplitude or direction at the associated internal key dose gradient feature point within a specific time window. This measure, called the displacement-dose coupling coefficient, is used to quantify the correlation strength between body surface motion and internal dose gradient perturbation at that spatial location. The calculation process is repeated at multiple sampling time points and multiple spatial location points.

[0027] Because the spatial distribution of key dose gradient features within a patient's body is non-uniform, their impact on the accuracy of the overall dose distribution varies. Therefore, the calculated raw displacement-dose coupling coefficient is spatially weighted. The weighting depends on the spatial distribution density map of the key dose gradient features. This density map is generated by estimating the spatial kernel density of all extracted key dose gradient feature points, reflecting the density of key dose gradient feature points per unit volume. Higher weights are assigned to high-density regions, such as areas with multiple dose hotspots or steep dose drop zones, or areas near highly radiation-sensitive organs; conversely, lower weights are assigned to low-density regions. The spatially weighted displacement-dose coupling coefficient contains information on the spatial differences in the influence of surface motion on the internal dose gradient.

[0028] To more accurately characterize the actual movement patterns of internal organs, it is necessary to map the spatially weighted displacement-dose coupling coefficients to a dedicated coordinate system that reflects the typical movement trajectory of the organs—the organ movement trajectory coordinate system. This coordinate system is typically established based on the patient's pre-treatment four-dimensional imaging data or organ movement prediction results based on biomechanical models. It defines the typical path of organ changes in three-dimensional space with the physiological cycle. Through a spatial transformation function, the positional information of the spatially weighted displacement-dose coupling coefficients in the original imaging coordinate system is transformed to the organ movement trajectory coordinate system. Within the organ movement trajectory coordinate system, spatial interpolation techniques, such as radial basis function interpolation or Kriging interpolation, are used to fit discrete sample points of the spatially weighted displacement-dose coupling coefficients mapped to this coordinate system. The interpolation process generates a continuous three-dimensional vector field covering the target organ or region—the three-dimensional anatomical displacement vector field. Each vector in this field not only describes the direction of displacement of the predicted anatomical structure in three-dimensional space at the corresponding spatial location but also quantifies the magnitude of the displacement, directly relating to the potential positional changes of internal organs caused by surface movement.

[0029] The constructed 3D dose flux distribution map and anatomical displacement feature set are input into a pre-built multimodal dose prediction model for processing. This model consists of three main functional modules: an image encoder, a motion feature encoder, and a multimodal fusion decoder. The image encoder is responsible for processing the 3D dose flux distribution map input. This encoder typically employs a 3D convolutional neural network structure, containing multiple 3D convolutional layers, pooling layers, and activation function layers. The 3D dose flux distribution map serves as the input tensor, undergoing progressive processing through these layers. The 3D convolutional layers slide 3D convolutional kernels across the dose distribution map, extracting local spatial pattern features at different scales. The pooling layers reduce the spatial dimensionality of the feature map, increasing the translation invariance and receptive field of the features. After multiple layers of processing, the image encoder finally outputs a fixed-length high-dimensional vector, called the dose flux feature vector. This vector encapsulates the key spatial distribution patterns and dosimetric characteristics of the input 3D dose flux distribution map.

[0030] The motion feature encoder is responsible for processing the anatomical displacement feature set as input. The anatomical displacement feature set is a dataset containing a time-series structure, with each time point containing key dose gradient features and predicted anatomical displacement vector field information. Motion feature encoders typically employ recurrent neural network (RNN) or temporal convolutional network (TCNN) architectures, specifically designed for processing sequential data. For each time step, the encoder receives the corresponding anatomical displacement feature data packet. RNNs utilize their hidden states to convey historical information, capturing long-term temporal dependencies in the sequence data. TCNNs extract features in the temporal dimension through one-dimensional causal convolution. The motion feature encoder processes the input sequence step-by-step, and its final output is a high-dimensional vector called the spatiotemporal motion feature vector. This vector encodes the time-evolving motion patterns and spatiotemporal correlation information inherent in the anatomical displacement feature set.

[0031] After being output from the image encoder and motion feature vector, respectively, the dose flux feature vector and spatiotemporal motion feature vector are fed into the multimodal fusion decoder. The core component of the multimodal fusion decoder is a fusion layer based on a cross-modal attention mechanism. This mechanism aims to dynamically integrate information from two different modalities. The cross-modal attention mechanism first calculates the mutual attention weights between the dose flux feature vector and the spatiotemporal motion feature vector. The calculation involves mapping the two feature vectors to the spaces of query vectors, key vectors, and value vectors, respectively. By calculating the similarity scores between the query vector and all key vectors, and then normalizing the results, the attention weight distribution is obtained. These weights indicate which features from the other modality are most relevant and important for feature prediction in the current modality during the fusion process. Based on the calculated attention weights, the value vectors are weighted and summed to achieve dynamic fusion of the feature information from the two modalities. The fused feature vector contains an information representation enhanced by cross-modal interaction. The fused feature vector then passes through other processing layers of the decoder, which may include fully connected layers or upsampling layers. Finally, the multimodal fusion decoder outputs a prediction result on a three-dimensional spatial grid, which is the first prediction result of the dose distribution deviation. This prediction result quantifies the amount of prediction deviation of the dose distribution at each location in three-dimensional space relative to the original plan, due to factors such as anatomical motion.

[0032] Example 3: See Figure 4 On a three-dimensional dose flux map, anatomical structures requiring special attention are identified—high-risk organs that are sensitive to radiation and where overexposure could lead to serious clinical consequences. These organs include, but are not limited to, the spinal cord, optic nerve pathways, brainstem, parotid gland, lungs, heart, liver, kidneys, rectum, bladder, and small intestine. The identification process utilizes predefined anatomical contours of high-risk organs. By overlaying these contours onto the three-dimensional dose flux map, dose distribution regions falling within the geometric space of the high-risk organs are located. A dose threshold is set to screen for high-dose areas of concern; this threshold can be determined based on the organ's known tolerable dose, a specific percentage of the prescribed dose, or the statistical characteristics of the dose distribution. Spatial areas where dose values ​​exceed the set threshold are marked as high-risk organ dose hotspots. These hotspots represent locations within or near the boundaries of high-risk organs where there is a potential risk of overexposure. The identification results are output as a hotspot map containing spatial location information and the dose values ​​exceeding the threshold.

[0033] For each identified high-risk organ dose hotspot distribution region, a dose-volume histogram is calculated. The dose-volume histogram describes the percentage of volume within that organ that received a dose equal to or higher than a specific dose value. The morphological characteristics of the generated dose-volume histogram curve in the high-dose region are analyzed. Particular attention is paid to the rate of decrease in the curve with increasing dose value. The dose-volume histogram curve in the high-dose region is mathematically processed to quantify the steepness of its change. The average rate of change of the curve's slope within a specified dose interval is calculated. This interval is typically set at the upper limit near the hotspot dose threshold. The average rate of change of the slope is obtained by calculating the slope of the tangent lines to the dose-volume histogram curve at the start and end points of this interval, taking the difference, and dividing by the length of the dose interval. This value reflects the severity of the rapid volume decrease caused by a small increase in dose accumulation in the core region of the dose hotspot. A larger average rate of change of the slope indicates that the dose distribution is highly concentrated in the high-dose region, with a small volume area receiving an extremely high dose; a smaller rate of change indicates that the high-dose region is relatively diffuse.

[0034] For specific isodose lines associated with high-risk organs in a three-dimensional dose flux distribution map, their geometric features are extracted. An isodose line is a surface formed by a set of points in space with the same dose value. Isodose lines meaningful for assessing organ risk are selected, such as isodose surfaces that are close to or exceed the organ's tolerance dose. On the selected isodose surfaces, characteristic quantities characterizing their local curvature are calculated. The curvature of an isodose line describes the degree to which the surface deviates from its tangent plane at a given point, and is a geometric index quantifying the steepness of spatial changes in dose distribution. At each point on the isodose surface, its local curvature value is calculated. This value is obtained by analyzing the differential geometric properties of the isodose surface at that point. A commonly used index for quantifying the local curvature of an isodose line surface is the mean curvature. The mean curvature value at each point on the isodose line is calculated, and statistical characteristic values ​​of the mean curvature of all points within the region enclosed by the isodose line are statistically analyzed, such as the average or maximum value. The mean curvature characteristic value reflects the clustering or dispersion trend of the dose value represented by the isodose line in spatial distribution. Regions with a large degree of curvature usually correspond to locations with larger dose gradients. The formula is as follows:

[0035] in: The curvature scalar value representing a point on the isodose line is a dimensionless quantity used to quantify the degree of local curvature at that point; T represents the unit tangent vector of the isodose line at that point, indicating the instantaneous direction of the curve; s represents the arc length parameter measured along the isodose line, in millimeters. The derivative vector of the unit tangent vector T as a function of arc length s, pointing towards the center of curvature of the isodose line, has a magnitude equal to the curvature. The value. The larger the value, the more the isodose line bends at that point, and the more drastic the spatial variation of the dose distribution is at that location.

[0036] Within the geometric boundary region of a high-risk organ, the spatial rate of dose change from a relatively high-value region inside the organ to a low-value region in normal tissue outside the organ is calculated. This rate is called the dose drop gradient. The magnitude of the gradient as the dose value decreases with increasing spatial distance in a direction perpendicular to the high-risk organ boundary is calculated. At each point on the organ boundary, the normal direction at that point is determined. Along this normal direction, the spatial derivative of the dose distribution, i.e., the rate of dose change with distance, is calculated. The magnitude of this gradient is the dose drop gradient. The characteristic values ​​of the dose drop gradient reflect the rate of formation of a dose protection zone at the edge of a high-risk organ. A steep drop gradient means that the dose decreases rapidly after leaving the organ boundary, helping to protect the normal tissue adjacent to the organ; a gentle drop gradient indicates that the high-dose region may extend further into normal tissue. For each high-risk organ, statistical characteristic values ​​of the dose drop gradient at its boundary are calculated, such as the average or minimum value.

[0037] A series of characteristic parameters calculated for each high-risk organ are combined: the average slope change rate of the dose accumulation curve in the high-dose region calculated for the dose hotspot distribution identified for that organ; the isodose curvature characteristic value of the isodose line covering the key dose levels of that organ; and the dose drop gradient characteristic value calculated along the boundary of that organ. These three features together form a vector representing the radiation biological characteristic parameters of that high-risk organ under a specific three-dimensional dose flux distribution map. The above process is repeated for all high-risk organs that need to be evaluated, and the results are summarized to obtain the set of radiation biological characteristic parameters.

[0038] The pre-defined radiation damage threshold model includes a structured database called the organ tolerance dose database. This database stores standardized information based on extensive clinical observations, biological dose-response models, and literature. For each high-risk organ type, the database contains multiple records, each defining the probability of that organ developing a specific type and severity of radiation damage at a specific dose level. The database also stores typical dose distribution feature pattern vectors associated with the damage probability. These feature pattern vectors contain components corresponding to the aforementioned extracted feature parameters, such as the rate of change of the slope of the typical dose cumulative curve at a specific dose level, the characteristic value of the curvature of the typical isodose line, and the characteristic value of the typical dose drop gradient.

[0039] The radiation biology feature parameter set extracted from the three-dimensional dose flux distribution map of the current treatment plan is matched with the feature pattern vectors of the corresponding organ types stored in the organ tolerance dose database. The matching process calculates the similarity between the current feature parameter vector and the reference feature pattern vector in the database. Various metrics can be used for similarity calculation, such as Euclidean distance, Manhattan distance, or cosine similarity. A smaller distance value or a higher cosine similarity indicates a greater similarity between the current feature parameter set and the feature patterns in the database. For each high-risk organ in the current plan, based on the calculated similarity value with the feature pattern vectors of multiple records in the database, and combined with the damage probability values ​​stored in these records, a weighted average or interpolation algorithm is used to calculate an estimated probability of radiation damage to that organ under the current dose distribution. This probability value is a value between zero and one.

[0040] A pre-defined dose safety assessment rule defines an acceptable radiation damage probability safety threshold for each high-risk organ. This threshold is typically set based on clinical consensus and treatment goals. The calculated current radiation damage probability estimates for each high-risk organ are compared to their corresponding safety thresholds. The degree to which the probability estimate for each organ deviates from its safety threshold is assessed. The degree of deviation can be quantified by absolute difference or relative proportion. Based on this deviation, a second prediction result for dose safety assessment is generated. The second prediction result can be represented in a tiered format, for example: if the probability estimate is below the safety threshold, it is marked as "low risk"; if it is close to the safety threshold but within an acceptable range, it is marked as "attention risk"; if it exceeds the safety threshold but within a certain tolerance, it is marked as "medium risk"; if it significantly exceeds the safety threshold, it is marked as "high risk". The result can also be output as a continuous risk score. The second prediction result clearly indicates the radiation safety status of each high-risk organ under the current plan.

[0041] Example 4: During stereotactic radiotherapy for lung cancer patients, the multimodal dose prediction model outputs a first prediction result. This result is presented in the form of a three-dimensional matrix, representing the deviation between the planned dose distribution and the actual predicted dose distribution. The model's own confidence assessment information is extracted from this prediction result. This confidence level originates from the model's internal uncertainty quantification module, reflecting the reliability of the model's prediction deviation for each spatial location. Assuming that at a specific optimization time point, for the dose deviation prediction of the spinal cord, an organ at risk, the model outputs a confidence score of 0.92 (range 0-1, where 1 represents the highest confidence). This score serves as the first weighting factor reflecting the reliability of the dose deviation prediction for the spinal cord region. Simultaneously, the radiation damage threshold model outputs a second prediction result, assessing the damage risk level of each high-risk organ under the current plan. This model calculates a radiation damage probability estimate of 0.15 for the spinal cord. According to preset rules, probability values ​​in the range of 0.1 to 0.2 are classified as "medium risk." Similarly, the estimated probability of lung tissue damage was 0.08, categorized as "low risk"; the estimated probability of esophageal damage was 0.25, categorized as "high risk". These risk levels were converted into numerical weights: low risk corresponds to a value of 1, medium risk to a value of 2, and high risk to a value of 3. Therefore, the second weighting factor for the spinal cord was 2, for the lungs 1, and for the esophagus 3. See Table 1, which shows the first and second weighting factors for different organs at this time point, along with their nonlinear normalization process.

[0042] Table 1: Data table for dynamic optimization weight calculation.

[0043]

[0044] Nonlinear normalization is achieved using the Sigmoid function transformation. Specifically, the product of the first and second weighting factors is input into the Sigmoid function for scaling. The parameters of the Sigmoid function are set to amplify the weights of high-risk or high-confidence organs. The processed normalized weight values ​​range from 0 to 1, summing to 1. As shown in the table, the esophagus, due to its high-risk level, receives the highest normalized weight of 0.89, followed by the spinal cord at 0.78, while the lung tissue, due to its low-risk level, receives the lowest weight of 0.37. These normalized weights represent the dynamic optimized weights for each organ, quantifying their protection priority within the current optimization cycle.

[0045] The treatment system monitors the anatomical displacement feature set data stream in real time. This feature set contains the predicted displacement vector field of internal organs derived from body surface motion. During monitoring, it was found that the predicted displacement amplitude of the esophagus in a certain direction exceeded the preset displacement safety threshold of 5 mm for three consecutive sampling cycles. The system immediately triggered a multi-level dose compensation strategy. The core basis of this strategy is the aforementioned dynamically optimized weights. Based on the monitored esophageal displacement amplitude, it was determined that it fell into the preset compensation level two range (displacement threshold of 5 mm to 8 mm). The dose gradient compensation layer corresponding to this level was activated. The dose gradient compensation layer is a predefined or real-time calculated spatial dose adjustment template. Its design goal is to compensate for the dose distribution distortion caused by the change in organ position when organ displacement occurs by increasing or decreasing the dose projection intensity in a specific spatial region, thereby maintaining target area dose coverage as much as possible and protecting organs at risk.

[0046] The compensation intensity parameter controls the magnitude of dose adjustment. This parameter is directly adjusted by the dynamic optimization weights. Since the dynamic optimization weight for the esophagus is as high as 0.89, its corresponding compensation intensity parameter is set to a high level. This means that the magnitude of dose adjustment will be significantly amplified in the compensation region along the predicted displacement direction of the esophagus. The specific value of the compensation intensity parameter is determined jointly by the compensation level and the dynamic optimization weights through a lookup table or a linear interpolation function. The compensation intensity parameter is then passed to the dose calculation core.

[0047] Based on the dose gradient compensation layer corresponding to the activated compensation level two, and the high compensation intensity parameter driven by the high weight of the esophagus, the system recalculates the configuration of the treatment beam. The optimization algorithm adjusts the beam angle weight distribution, increasing the contribution weight of beams incident from specific angles to the compensation region, and reducing the beam angle weights that may cause high doses to the displaced esophagus. The goal of reconstructing the beam angle weight distribution is to create a dose trough around the new position of the esophagus, while maintaining the dose uniformity of the target area as much as possible.

[0048] Based on the reconstructed beam angle weight distribution and the requirements of the dose gradient compensation layer, the blade motion trajectory parameters of the multi-leaf collimator are adjusted. The motion trajectory of the multi-leaf collimator blades is recalculated to ensure that the opening shape formed by the blades precisely matches the shape of the compensated radiation field. At the blade position corresponding to the predicted displacement region of the esophagus, the blade motion speed is adjusted to control the opening and closing time of that region, achieving precise local dose control. Simultaneously, the dose rate control curve is dynamically adjusted. When the blade moves to a position requiring enhanced compensation, the dose rate is increased accordingly to rapidly deliver the compensated dose; when the blade moves to a shielding position requiring esophageal protection, the dose rate is reduced or even paused to avoid unnecessary irradiation. The close coordination between the adjustment of the blade motion trajectory parameters and the dose rate control curve ensures that the spatial dose adjustment requirements defined by the dose gradient compensation layer can be accurately achieved at the physical execution level.

[0049] Example 5: During intensity-modulated radiotherapy (IMRT) for prostate cancer patients, the system dynamically optimized the leaf trajectory parameters of the multi-leaf collimator and the dose rate control curve based on weighted adjustments. As treatment continued, the system simultaneously performed an automated assessment of treatment risks. This assessment was based on real-time transmitted dose images acquired by the electronic field imaging device in the treatment room, real-time readings from the implanted semiconductor dosimeter, and a continuously updated anatomical displacement feature set data stream. The assessment algorithm compared the real-time dose monitoring data with the three-dimensional dose flux distribution map set in the original treatment plan, analyzing dose differences voxel by voxel. The algorithm identified a specific area on the anterior rectal wall near the prostate target area where the real-time monitored dose values ​​at multiple consecutive sampling points were significantly higher than the planned dose values, exceeding the preset deviation tolerance range. This area was automatically marked as a dose hotspot distribution risk marker, indicating a potential risk of over-irradiation. Simultaneously, the continuously input anatomical displacement feature set was analyzed, particularly the predicted rectal displacement vector field data. The algorithm detected that during specific treatment periods, the predicted rectal displacement amplitude along the anterior-posterior direction repeatedly approached the preset displacement safety threshold, triggering a displacement risk warning. This time period and its corresponding spatial location were marked as organ displacement risk markers, indicating that the rectum was at risk of positional instability during this period.

[0050] The system automatically summarizes the above identification results and generates a structured list of treatment execution risk characteristics. This list uses a standardized data format: each identified risk event is assigned a unique event identifier; the precise timestamp of the event is recorded; the corresponding three-dimensional spatial coordinate range is recorded; the event type is clearly marked (e.g., "dose hotspot" or "organ displacement"); key quantitative indicators of the trigger marker are recorded (e.g., percentage of dose deviation, displacement amplitude); the dynamic optimization weight value corresponding to the associated event; the first and second prediction results output by the multimodal dose prediction model input when the associated event occurs; the treatment parameter adjustment records activated when the associated event occurs (e.g., specific leaf trajectory correction parameters, dose rate adjustment values); and the version of the three-dimensional dose flux distribution map and the snapshot of the anatomical displacement feature set used when the associated event occurs. For the aforementioned identified rectal region dose hotspot risk events and displacement risk events, the list records in detail all associated data items, including time, spatial location, dose deviation, predicted displacement, and the optimization weights applied at the time. This risk characteristic list is not isolated; it is bound to a complete snapshot of the system's state at the time the event is triggered.

[0051] This list of treatment execution risk characteristics is considered a valuable data sample reflecting complex risk scenarios in actual treatment execution. The system automatically adds this list and its associated complete status data (including the associated 3D dose-flux distribution map, anatomical displacement feature set, first prediction result, second prediction result, dynamically optimized weight values, and finally adjusted treatment parameters) as a new, structured training sample to the historical dataset used to train the multimodal dose prediction model. The original historical dataset mainly contains simulation data from the treatment planning phase and some historical execution data. The newly added data sample provides specific scenario data under real treatment execution conditions when significant organ displacement occurs, leading to actual monitored dose distribution deviations, enriching the dataset's diversity and representativeness in risk scenarios. The data addition process follows strict version control and data verification procedures to ensure the integrity of the new data and compatibility with existing data formats.

[0052] After obtaining an updated training dataset containing newly added risk feature samples, the system initiates a retraining process for the multimodal dose prediction model. The retraining focuses on optimizing specific parameters in the model's core component—the multimodal fusion decoder—namely, the fusion decision parameters. These parameters primarily reside in the cross-modal attention mechanism layer responsible for integrating dose flux feature vectors (from the image encoder) and spatiotemporal motion feature vectors (from the motion feature encoder). They control how attention weights are calculated, determining the degree of mutual influence between the two modalities and the areas of focus during the fusion process. During retraining, the model iteratively learns using the updated training dataset. The optimization algorithm calculates the difference between the model's predicted dose distribution deviation and the actual recorded or monitored dose deviation in the samples, based on the input samples, including newly added samples containing dose hotspots and organ displacement risk markers. Based on this difference, the gradient of the fusion decision parameters is calculated using backpropagation, and the values ​​of these parameters are adjusted. The training objective is to minimize the model's prediction error across the entire training set containing risk scenario samples. Specifically, when the input samples contain features such as newly added rectal displacement leading to hotspot risk, the optimization process significantly adjusts the fusion decision parameters. This allows the cross-modal attention mechanism to more accurately capture the spatiotemporal correlation of such organ displacement and dose hotspot association patterns, thereby outputting more accurate dose distribution bias predictions when encountering similar patterns in the future. Parameter tuning is a continuous numerical optimization process that does not change the basic architecture of the model.

[0053] After model retraining, a new version of model parameters is generated. The system automatically deploys this updated version of the multimodal dose prediction model to the online treatment environment, replacing the original model version. The deployment process ensures a seamless switch and does not affect ongoing treatment. In subsequent treatment, the system uses the updated model for real-time dose prediction and optimization decisions. When encountering patterns similar to previously recorded rectal displacement and hotspot risks, the updated model, with its optimized fusion decision parameters, can theoretically identify such associations more sensitively and predict the dose deviation risk in the rectal region earlier and more accurately. This improved predictive ability allows the system to generate corresponding dynamic optimization weights more promptly and adjust treatment parameters more effectively, such as more precisely controlling the blade shielding or dose rate limitation in the rectal region to avoid or mitigate the risk of actual dose hotspot formation. The entire process constitutes a closed-loop self-optimizing system: risks are identified and data is collected during execution; the data is used to update the model training set; the model fine-tunes parameters for risk patterns; and the model is updated to improve future risk handling capabilities.

[0054] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0055] 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 interactive radiotherapy planning system optimization method, characterized by, include: Acquire the patient's medical imaging data and real-time body surface motion monitoring data, construct a three-dimensional dose-flux distribution map based on the medical imaging data, and generate an anatomical displacement feature set based on the real-time body surface motion monitoring data; The three-dimensional dose flux distribution map and the anatomical displacement feature set are input into a pre-constructed multimodal dose prediction model for processing to obtain the first prediction result of the dose distribution deviation; Radiation biological characteristic parameters are extracted from the three-dimensional dose flux distribution map and matched with a preset radiation damage threshold model to obtain a second prediction result for dose safety assessment. Dynamic optimization weights are generated based on the first prediction result and the second prediction result. Treatment plan parameters are adjusted according to the dynamic optimization weights to achieve radiotherapy optimization. The multimodal dose prediction model includes an image encoder, a motion feature encoder, and a multimodal fusion decoder; The step of inputting the three-dimensional dose flux distribution map and the anatomical displacement feature set into a pre-constructed multimodal dose prediction model for processing includes: The dose flux feature vector of the three-dimensional dose flux distribution map is extracted by the image encoder; The spatiotemporal motion feature vector of the anatomical displacement feature set is parsed by the motion feature encoder. The dose flux feature vector and the spatiotemporal motion feature vector are input into the multimodal fusion decoder based on the cross-modal attention mechanism, and the first prediction result of the dose distribution deviation is output. The extraction of radiation biological characteristic parameters based on the three-dimensional dose flux distribution map includes: Identify the dose hotspot distribution of high-risk organs in the three-dimensional dose flux distribution map; Calculate the rate of change of the slope of the cumulative dose curve for the dose hotspot distribution of the high-risk organs; Extract isodose curvature features and dose drop gradient features of dose-sensitive organs; The rate of change of the slope of the dose accumulation curve, the curvature characteristics of the isodose line, and the dose drop gradient characteristics are combined into a set of radiation biological characteristic parameters.

2. The method of claim 1, wherein, The step of constructing a three-dimensional dose-flux distribution map based on the medical imaging data and generating an anatomical displacement feature set based on the real-time body surface motion monitoring data includes: The medical image data is used to perform dose flux simulation calculations to generate an initial three-dimensional dose flux distribution map, and key dose gradient features are extracted from the initial three-dimensional dose flux distribution map; The spatiotemporal correlation between the real-time body surface motion monitoring data and the key dose gradient features is quantified to generate an anatomical displacement vector field; The key dose gradient features and the anatomical displacement vector field are standardized and integrated in a time series format to generate the anatomical displacement feature set.

3. The method for optimizing an interactive radiotherapy planning system according to claim 2, characterized in that, The quantification of the spatiotemporal correlation between the real-time body surface motion monitoring data and the key dose gradient features includes: Phase alignment analysis was performed on the synchronously acquired real-time body surface motion monitoring data and the key dose gradient features to calculate the displacement-dose coupling coefficient; Based on the distribution density of the key dose gradient features, the displacement-dose coupling coefficient is spatially weighted. The weighted displacement-dose coupling coefficients are mapped to the organ motion trajectory coordinate system to generate a three-dimensional anatomical displacement vector field.

4. The method for optimizing an interactive radiotherapy planning system according to claim 3, characterized in that, The matching with the preset radiation damage threshold model includes: The radiation biology feature parameter set is matched with a preset organ tolerance dose database for feature similarity. The probability value of radiation damage to each high-risk organ is calculated based on the matching similarity. The second prediction result is generated based on the degree of deviation between the radiation damage probability value and the preset safety threshold.

5. The method of claim 1, wherein, The step of generating dynamic optimization weights based on the first prediction result and the second prediction result includes: The dose deviation confidence level in the first prediction result is extracted as the first weighting factor; The organ damage risk level in the second prediction result is used as the second weighting factor; The first weighting factor and the second weighting factor are subjected to nonlinear normalization to generate dynamic optimized weights for organ protection priority.

6. The method of claim 5, wherein the optimization is performed by a system comprising: a user interface; a processor; and a memory coupled to the processor and storing computer readable instructions that, when executed by the processor, cause the processor to perform the optimization. The step of adjusting the treatment plan parameters according to the dynamically optimized weights includes: When the anatomical displacement feature set shows that the organ displacement exceeds the threshold, a multi-level dose compensation strategy based on the dynamic optimization weight is activated. The beam angle weight distribution is reconstructed based on the multi-level dose compensation strategy; Adjust the blade motion trajectory parameters and dose rate control curve of the multi-leaf collimator.

7. The method for optimizing an interactive radiotherapy planning system according to claim 6, characterized in that, The activation of the multi-level dose compensation strategy based on the dynamically optimized weights includes: Based on the organ displacement amplitude, compensation levels are divided, and a multi-level dose gradient compensation layer is constructed. Obtain the deviation between real-time dose monitoring data and the planned dose; The compensation intensity parameters of the multi-level dose gradient compensation layer are dynamically adjusted based on the deviation.

8. The method for optimizing an interactive radiotherapy planning system according to claim 7, characterized in that, After adjusting the blade motion trajectory parameters and dose rate control curve of the multi-leaf collimator, the following is also included: Generate a treatment execution risk feature list, which includes organ displacement risk markers and dose hot and cold spot distribution markers; The training dataset of the multimodal dose prediction model is updated according to the risk feature list; The fusion decision parameters of the multimodal fusion decoder are iteratively optimized based on the updated training dataset.

Images