Self-adaptive radiotherapy starting opportunity prediction method and system

Abstract

The invention relates to the technical field of medical image analysis, and provides a self-adaptive radiotherapy starting opportunity prediction method and system, and the method comprises the steps: obtaining a quantitative ART index of a radiotherapy starting condition through image analysis and calculation; utilizing an image omics tool to obtain image omics characteristics of the graded image target region and the key organ at risk; aRT indexes and radiomics features form a feature set, and recursive feature elimination is combined with KL divergence to perform feature screening on the feature set; dividing the screened multi-dimensional features into a training set and a test set, and training a bidirectional LSTM and attention mechanism combined model through the training set to form a space-time fusion prediction model; and inputting the test set into the space-time fusion prediction model, and outputting a radiotherapy starting opportunity prediction result. According to the method, targeted self-adaptive starting opportunity aid decision-making which varies from person to person can be provided for each patient, and the treatment accuracy and quality of different patients can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, and in particular to an adaptive radiotherapy initiation timing prediction method and system. Background Technology

[0002] Most existing medical linear accelerators employ image-guided radiation therapy (IGRT) using fractionated images (including CBCT, MR, FBCT, etc.). This involves tracking anatomical changes in the patient by scanning fractionated images online before each fraction. If the patient's original treatment plan is deemed to deviate too far from the treatment goal, adjustments are necessary; this process is called online adaptive radiotherapy (ART-online). Adaptive radiotherapy aims to rapidly provide clinicians with feedback before or during treatment, allowing them to adjust the treatment plan based on changes in physiological tissues and patient positioning.

[0003] The process of adaptive radiotherapy involves acquiring two-dimensional or three-dimensional images of the patient at each fraction of treatment from the start of the course, and measuring the positioning error offline. Based on the initial measurements (5-9 fractions), the positioning error for the entire treatment course is predicted. The spacing between the PTV and CTV is then adjusted accordingly, the treatment plan is modified, and subsequent fractions are administered according to the revised plan. Beyond adjusting the spacing based on individual positioning errors, adaptive radiotherapy has recently been extended to higher levels, such as adjusting the dose of subsequent fractions based on the cumulative actual radiation dose received by the patient in each fraction, or adjusting the target area and / or prescribed dose based on the tumor's response to treatment during the course.

[0004] Currently, traditional adaptive radiotherapy technology has the following three major drawbacks: Resource inefficiency: The current ART process is time-consuming (patients wait an average of >30 minutes) and requires professional physicists to manually analyze images in segments.

[0005] Lack of decision-making criteria: There is a lack of quantitative indicators to determine when to initiate ART (such as a clear threshold like "PTV dose deviation ≥5%"). Therefore, current online adaptive systems execute adaptive planning for each fraction, which is time-consuming, labor-intensive, increases unnecessary workload, and delays timely treatment for patients.

[0006] Insufficient data utilization: Traditional methods ignore the dynamic changes in temporal radiomics features during treatment.

[0007] Moreover, although traditional adaptive radiotherapy has been proven to effectively improve the accuracy and quality of patient treatment, there are still some limitations that hinder its widespread application: (1) Currently, medical resources are limited, and patients have to wait a long time online for treatment, resulting in low ART efficiency and a time-consuming and labor-intensive process; (2) CBCT imaging range is limited and often cannot cover the entire treatment area, so redesigning radiotherapy plans with it will result in dose errors; (3) In order to avoid significantly increasing the additional radiation dose caused by CBCT and increasing radiation risks, the current clinical standard procedure does not scan CBCT images before each treatment fraction; (4) Based on the above three problems, the plan of performing ART in each fraction is too difficult for most radiotherapy units and patients. A more appropriate approach is to predict the timing of ART initiation in advance. However, there is currently no unified standard for how much a patient's radiotherapy plan needs to deviate from its expected target to trigger the use of adaptive radiotherapy. One consequence of the above limitations is that some patients do not receive the necessary and accurate plan adjustments, even though they could have benefited more from the treatment and improved the efficacy. Summary of the Invention

[0008] The purpose of this invention is to solve at least one technical problem in the background art and to provide an adaptive radiotherapy initiation timing prediction method and system.

[0009] To achieve the above objectives, the present invention provides an adaptive radiotherapy initiation timing prediction method, comprising: The ART index, which quantifies the conditions for initiating radiotherapy, was obtained through image analysis and calculation. Utilize radiomics tools to obtain radiomics characteristics of target areas and key organs at risk in fractionated imaging; The ART index and radiomics features are combined to form a feature set, and the recursive feature elimination combined with KL divergence is used to filter the features in the feature set; The multidimensional features obtained after screening are divided into training set and test set. The model combining bidirectional LSTM and attention mechanism is trained through the training set to form a spatiotemporal fusion prediction model. The test set is input into the spatiotemporal fusion prediction model, and the predicted timing of radiotherapy initiation is output.

[0010] According to one aspect of the invention, the ART metric includes an initial ART metric and a phased ART metric, both of which include: Tumor volume shrinkage rate; Planned target area dose deviation; The rate of overdose on organs at risk; Biological effects TCP and NTCP.

[0011] According to one aspect of the invention, the radiomics features include: Patient's clinical diagnostic characteristics; Radiomic features of the initial images; Initial dose-omics characteristics; Radiomics features of fractional images; fractional dose omics characteristics; Cumulative dose-omics characteristics.

[0012] According to one aspect of the present invention, the feature selection of the feature set using recursive feature elimination combined with KL divergence includes: (1) Calculate the delta eigenvalue and retain the eigenvalues ​​with time series fluctuations >30%; (2) Recursive feature elimination is used for initial screening, retaining the top 30% of features; (3) KL divergence time-series sensitivity enhancement screening: Among them, P t Let f be the characteristic distribution at time t, P0 be the initial distribution, and f i Let i be the i-th feature.

[0013] According to one aspect of the present invention, the structure of the model combining bidirectional LSTM and attention mechanism includes: Input layer: Receives multidimensional features, including spatial and temporal features; Bidirectional LSTM layer: It performs bidirectional modeling on the sequence of input multidimensional features, extracting time series features from both the forward and reverse directions; the forward LSTM captures the dependencies from the past to the present, and the reverse LSTM captures the dependencies from the future to the present, thus achieving comprehensive modeling of the time series. Attention Mechanism Layer: Based on the bidirectional LSTM output, an attention mechanism is introduced to perform weighted processing on the temporal steps and spatial features between different temporal features; Output layer: The feature vectors processed by the attention mechanism are fused to generate the final prediction result, which is the prediction value of whether or not radiotherapy will be initiated.

[0014] According to one aspect of the present invention, when the model fuses spatial features, it first assigns different attention weights to different spatial features through an attention mechanism layer to achieve selective fusion of key spatial features.

[0015] According to one aspect of the present invention, when fusing temporal features, the model first calculates the correlation between different time steps through an attention mechanism layer to capture global context information and generate a weighted temporal state feature representation; then, the temporal state features are used to perform sequence modeling through an LSTM layer, and a recursive gating mechanism is used to capture long sequences and long-term dependencies. The attention mechanism layer introduces an attention scoring mechanism, including: Calculate the energy value: Energy value e t Used to measure the current query vector h T With historical key vector h t The similarity between them is calculated using the following formula: ; Where v is the weight vector, W a and U a It is a learnable parameter matrix, b a It is a bias term; Normalized weights: The attention weight α is obtained by normalizing the energy value using the Softmax function. t To ensure that the sum of all attention weights is 1, the normalization formula is: ; Generate context vectors: Using attention weight α t For the historical key vector h t Perform a weighted summation to generate a context vector c, which incorporates the information most relevant to the current query: ; Input the context vector c into a single neuron layer activated by Sigmoid, and output the ART activation probability: ; When p ART Adaptive radiotherapy is triggered when the value is >0.8.

[0016] To achieve the above objectives, the present invention also provides an adaptive radiotherapy initiation timing prediction system, comprising: The ART index acquisition module calculates ART indexes that quantify the conditions for initiating radiotherapy through image analysis. The radiomics feature acquisition module uses radiomics tools to acquire radiomics features of target areas and key organs at risk in fractional imaging. The feature selection module combines ART indicators and radiomics features into a feature set, and uses recursive feature elimination combined with KL divergence to select features from the feature set. The spatiotemporal fusion prediction model construction module divides the selected multidimensional features into training and testing sets. The training set is used to train the model that combines bidirectional LSTM with attention mechanism to form a spatiotemporal fusion prediction model. The test set is input into the spatiotemporal fusion prediction model, and the predicted timing of radiotherapy initiation is output.

[0017] To achieve the above objectives, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the adaptive radiotherapy initiation timing prediction method as described above.

[0018] To achieve the above objectives, the present invention also provides a computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and the computer program, when executed by a processor, implements the adaptive radiotherapy initiation timing prediction method as described above.

[0019] According to the present invention, an adaptive radiotherapy initiation decision-making strategy is designed based on dose analysis and omics information from previous treatments. This strategy allows for the selection of an appropriate timing for initiating adaptive radiotherapy for individualized patients, fully leveraging the advantages of adaptive radiotherapy and further improving treatment outcomes. The present invention provides targeted, individualized adaptive initiation timing support for each patient, effectively improving the accuracy and quality of treatment for different patients. Attached Figure Description

[0020] Figure 1 The flowchart illustrates an adaptive radiotherapy initiation timing prediction method according to one embodiment of the present invention. Detailed Implementation

[0021] The invention will now be discussed with reference to exemplary embodiments. It should be understood that the described embodiments are merely intended to enable those skilled in the art to better understand and thus implement the invention, and are not intended to imply any limitation on the scope of the invention.

[0022] As used herein, the term "comprising" and its variations are to be interpreted as open-ended terms meaning "including but not limited to". The term "based on" is to be interpreted as "at least partially based on". The terms "one embodiment" and "an embodiment" are to be interpreted as "at least one embodiment".

[0023] Figure 1 The flowchart schematically illustrates an adaptive radiotherapy initiation timing prediction method according to one embodiment of the present invention. Figure 1 As shown, in this embodiment, the adaptive radiotherapy initiation timing prediction method includes: The ART index, which quantifies the conditions for initiating radiotherapy, was obtained through image analysis and calculation. Utilize radiomics tools to obtain radiomics characteristics of target areas and key organs at risk in fractionated imaging; The ART index and radiomics features are combined to form a feature set, and the recursive feature elimination combined with KL divergence is used to filter the features in the feature set; The multidimensional features obtained after screening are divided into training set and test set. The model combining bidirectional LSTM and attention mechanism is trained through the training set to form a spatiotemporal fusion prediction model. Input the test set into the spatiotemporal fusion prediction model and output the prediction results of the radiotherapy initiation timing.

[0024] Furthermore, according to one embodiment of the present invention, the ART metric includes an initial ART metric and a phased ART metric, both of which include: Tumor volume shrinkage rate; Planned target area dose deviation; The rate of overdose on organs at risk; Biological effects TCP and NTCP.

[0025] Furthermore, according to one embodiment of the present invention, radiomics features include: Patient's clinical diagnostic characteristics; Radiomic features of the initial images; Initial dose-omics characteristics; Radiomics features of fractional images; fractional dose omics characteristics; Cumulative dose-omics characteristics.

[0026] In this embodiment, the patient's clinical diagnostic characteristics are: clinical indicators and diagnostic results obtained during the patient's diagnosis process, including age, gender, weight, BMI, TNM stage, presence or absence of paraneoplastic syndromes, radiotherapy and chemotherapy status, tumor location, and pathological type. Radiomics features of the initial image: Volume, shape, texture, grayscale, and high-order features extracted from the CTV region after CTV delineation of the scanned localization CT image; Initial dose-omics features: texture, grayscale, and high-order features extracted from the CTV region in the initial 3D dose distribution cloud map generated based on localization CT. Radiomics features of fractional images: Enhance the scanned fractional images and delineate the CTV (Central Television Volume). Extract volume, shape, texture, grayscale, and high-order features from the delineated CTV region. Fractional dose profiling features: After enhancing and deformably registering the scanned fractional images to the initial image, dose calculation is performed to obtain fractional three-dimensional dose distribution cloud maps. Texture, grayscale, and high-order features are extracted from the CTV region in the dose cloud maps. Cumulative dose profiling features: all fractional 3D dose distribution cloud maps are accumulated point-to-point, and texture, grayscale, and high-order features are extracted from the CTV region in the accumulated dose cloud map.

[0027] In this embodiment, the ART index is obtained by acquiring basic feature data through, for example, CT and CBCT, and then performing relevant calculations; the radiomics features are obtained through, for example, CT, CBCT, and spatial metrology distribution.

[0028] In this embodiment, ART indicators are constructed based on quantitative indicators such as tumor status, dose analysis, and biological effects. By comparing the initial ART indicator and the fractional ART indicator, and combining manual analysis for confirmation, the ART indicator for each fraction is determined to be 0 or 1.

[0029] Furthermore, because radiomics tools are used to extract radiomics features from fractionated images of the target region and critical organs at risk, typically hundreds of features are obtained from a single labeled region. The target region and adjacent organs at risk require feature collection from multiple labeled regions. Since the number of features at a single moment is substantial, and radiomics features from each fractionated image before the determined adaptive initiation time also need to be considered, the sheer number of features is unsustainable and hinders the training of the decision model. Therefore, feature filtering is necessary to identify those features that change significantly at the adaptive initiation time, thus reducing the dimensionality of the decision model input to a reasonable range.

[0030] According to one embodiment of the present invention, feature filtering of a feature set is performed using recursive feature elimination combined with KL divergence, including: (1) Calculate the delta feature value (subsequent features minus initial features), and retain features with time-series fluctuations >30%; (2) Recursive feature elimination initial screening, retaining the top 30% of features. The top 30% of features refer to the features whose importance (such as absolute value of coefficients, feature importance of tree model) is ranked in the top 30% during the iteration process. (3) KL divergence time-series sensitivity enhancement screening (threshold > 0.3): ; Among them, P t Let f be the characteristic distribution at time t, P0 be the initial distribution, and f i Let i be the i-th feature.

[0031] Furthermore, according to one embodiment of the present invention, the structure of the model combining bidirectional LSTM and attention mechanism includes: Input layer: Receives multidimensional features, including spatial and temporal features; Bidirectional LSTM layer: It performs bidirectional modeling on the sequence of input multidimensional features, extracting time series features from both the forward and reverse directions; the forward LSTM captures the dependencies from the past to the present, and the reverse LSTM captures the dependencies from the future to the present, thus achieving comprehensive modeling of the time series. Among them, forward LSTM: encodes the historical evolution from the initial time (t=0) to the current time (t=T); ; Reverse LSTM: Backtracks from the current time (t=T) to the initial time (t=0) to capture inverse dependencies; ; Final hidden state: ; d represents the number of hidden units in a single LSTM layer; in this implementation, d = 256. The LSTM gating update mechanism uses a forget gate to determine which historical information has been lost, an input gate to filter out valuable new information and update the state, and an output gate to control the current hidden state. When a segment of an image is of poor quality due to motion artifacts, the forget gate can reduce its weight, while the input gate can enhance features such as tumor volume mutations.

[0032] Attention Mechanism Layer: Based on the bidirectional LSTM output, an attention mechanism is introduced to perform weighted processing on the temporal steps and spatial features between different temporal features; Output layer: The feature vectors processed by the attention mechanism are fused to generate the final prediction result, which is the prediction value of whether or not radiotherapy will be initiated.

[0033] In this embodiment, the adaptive radiotherapy (ART) initiation timing prediction task requires processing two types of key information simultaneously: spatial features and temporal features. Spatial features include changes in the anatomical structure of the target area and organs at risk in each image (such as volume, shape, and texture), while temporal features involve dynamic evolution patterns such as dose accumulation and tumor regression during treatment. Traditional RNNs or single-time-series models have the following drawbacks: unidirectional LSTMs can only capture unidirectional dependencies from history to the present, ignoring potential future trends; without an attention mechanism, the model treats all time steps equally, failing to focus on key treatment nodes (such as sudden tumor regression).

[0034] Based on the analysis of the adaptive initiation timing prediction task, the model not only needs to examine the characteristics of the current moment, but also needs to compare the differences and changes in radiomic features between the current moment and the past state, especially the initial plan, in order to make a reasonable estimate of the probability that adaptive radiotherapy needs to be initiated at the current moment. Therefore, this implementation proposes the aforementioned adaptive spatiotemporal fusion prediction model, which achieves multi-dimensional information fusion through bidirectional LSTM combined with an attention mechanism. This model can simultaneously capture the bidirectional dependencies of time series and use the attention mechanism to weight key time steps and spatial features, thereby improving prediction accuracy and model interpretability.

[0035] Furthermore, according to one embodiment of the present invention, when fusing spatial features, the model first assigns different attention weights to different spatial features (such as target volume, shape, texture, etc.) through an attention mechanism layer, thereby achieving selective fusion of key spatial features. This method not only improves the interpretability of the model but also enhances its ability to capture complex spatiotemporal dependencies.

[0036] Furthermore, according to one embodiment of the present invention, when the model fuses time features, it first calculates the correlation between different time steps through an attention mechanism layer to capture global context information and generate a weighted time state feature representation; then, the time state features are used to perform sequence modeling through an LSTM layer, and a recursive gating mechanism is used to capture long sequences and long-term dependencies, thereby obtaining a richer time feature representation. To address the issue of traditional LSTM treating all time steps "equally," an attention scoring mechanism is introduced into the attention mechanism layer, including: Calculate the energy value: Energy value e t Used to measure the current query vector h T With historical key vector h t The similarity between them is calculated using the following formula: ; Where v is the weight vector, W a and U a It is a learnable parameter matrix, b a It is a bias term; Normalized weights: The attention weight α is obtained by normalizing the energy value using the Softmax function. t This step ensures that the sum of all attention weights is 1, guaranteeing that the model focuses on the historical time step most relevant to the current query; the normalization formula is: ; Generate context vectors: Using attention weight αt For the historical key vector h t Perform a weighted summation to generate a context vector c, which incorporates the information most relevant to the current query: ; In this way, the model can dynamically adjust its focus on different time steps based on the context, thereby improving the accuracy of predictions.

[0037] Input the context vector c into a single neuron layer activated by Sigmoid (i.e., the output layer of the model), and output the ART activation probability: ; When p ART Adaptive radiotherapy is triggered when the value is >0.8.

[0038] Furthermore, according to one embodiment of the present invention, the method of the present invention can take the initial features (radiomics features of the initial image and initial doseomics features) as x0 input, take the fractionated image features (radiomics features of the fractionated image and fractionated doseomics features) at the current time of investigation as xt input, and take the average state avg([x1, x2, ..., xt-1]), the maximum state max([x1, x2, ..., xt-1]), and the minimum state min([x1, x2, ..., xt-1]) of all states {x1, x2, ..., xt-1} between x0 and xt as input, and output the probability of initiating adaptive radiotherapy at the current time after model processing. The state of non-adaptive initiation timing can be used as a negative sample, and the true value of the adaptive initiation probability at that time is set to 0; the state of adaptive initiation timing determined by dose analysis is used as a positive sample, and the true value of its corresponding adaptive initiation probability is set to 1. Based on the collected patient data, the method is trained based on the above-mentioned various feature inputs to provide a reference for adaptive initiation timing decision-making.

[0039] According to the above-described scheme of the present invention, an adaptive radiotherapy initiation decision strategy is designed based on dose analysis and omics information from previous treatments. This scheme allows for the selection of an appropriate timing for initiating adaptive radiotherapy for individualized patients, fully leveraging the advantages of adaptive radiotherapy and further improving treatment outcomes. The present invention can provide each patient with targeted, individualized adaptive initiation timing support, effectively improving the accuracy and quality of treatment for different patients.

[0040] Furthermore, to achieve the above objectives, the present invention also provides an adaptive radiotherapy initiation timing prediction system, comprising: The ART index acquisition module calculates ART indexes that quantify the conditions for initiating radiotherapy through image analysis. The radiomics feature acquisition module uses radiomics tools to acquire radiomics features of target areas and key organs at risk in fractional imaging. The feature selection module combines ART indicators and radiomics features into a feature set, and uses recursive feature elimination combined with KL divergence to select features from the feature set. The spatiotemporal fusion prediction model construction module divides the selected multidimensional features into training and testing sets. The training set is used to train the model that combines bidirectional LSTM with attention mechanism to form a spatiotemporal fusion prediction model. The test set is input into the spatiotemporal fusion prediction model, and the predicted timing of radiotherapy initiation is output.

[0041] Furthermore, according to one embodiment of the present invention, the ART metric includes an initial ART metric and a phased ART metric, both of which include: Tumor volume shrinkage rate; Planned target area dose deviation; The rate of overdose on organs at risk; Biological effects TCP and NTCP.

[0042] Furthermore, according to one embodiment of the present invention, radiomics features include: Patient's clinical diagnostic characteristics; Radiomic features of the initial images; Initial dose-omics characteristics; Radiomics features of fractional images; fractional dose omics characteristics; Cumulative dose-omics characteristics.

[0043] In this embodiment, the patient's clinical diagnostic characteristics are: clinical indicators and diagnostic results obtained during the patient's diagnosis process, including age, gender, weight, BMI, TNM stage, presence or absence of paraneoplastic syndromes, radiotherapy and chemotherapy status, tumor location, and pathological type. Radiomics features of the initial image: Volume, shape, texture, grayscale, and high-order features extracted from the CTV region after CTV delineation of the scanned localization CT image; Initial dose-omics features: texture, grayscale, and high-order features extracted from the CTV region in the initial 3D dose distribution cloud map generated based on localization CT. Radiomics features of fractional images: Enhance the scanned fractional images and delineate the CTV (Central Television Volume). Extract volume, shape, texture, grayscale, and high-order features from the delineated CTV region. Fractional dose profiling features: After enhancing and deformably registering the scanned fractional images to the initial image, dose calculation is performed to obtain fractional three-dimensional dose distribution cloud maps. Texture, grayscale, and high-order features are extracted from the CTV region in the dose cloud maps. Cumulative dose profiling features: all fractional 3D dose distribution cloud maps are accumulated point-to-point, and texture, grayscale, and high-order features are extracted from the CTV region in the accumulated dose cloud map.

[0044] In this embodiment, the ART index is obtained by acquiring basic feature data through, for example, CT and CBCT, and then performing relevant calculations; the radiomics features are obtained through, for example, CT, CBCT, and spatial metrology distribution.

[0045] In this embodiment, ART indicators are constructed based on quantitative indicators such as tumor status, dose analysis, and biological effects. By comparing the initial ART indicator and the fractional ART indicator, and combining manual analysis for confirmation, the ART indicator for each fraction is determined to be 0 or 1.

[0046] Furthermore, because radiomics tools are used to extract radiomics features from fractionated images of the target region and critical organs at risk, typically hundreds of features are obtained from a single labeled region. The target region and adjacent organs at risk require feature collection from multiple labeled regions. Since the number of features at a single moment is substantial, and radiomics features from each fractionated image before the determined adaptive initiation time also need to be considered, the sheer number of features is unsustainable and hinders the training of the decision model. Therefore, feature filtering is necessary to identify those features that change significantly at the adaptive initiation time, thus reducing the dimensionality of the decision model input to a reasonable range.

[0047] According to one embodiment of the present invention, feature filtering of a feature set is performed using recursive feature elimination combined with KL divergence, including: (1) Calculate the delta feature value (subsequent features minus initial features), and retain features with time-series fluctuations >30%; (2) Recursive feature elimination initial screening, retaining the top 30% of features. The top 30% of features refer to the features whose importance (such as absolute value of coefficients, feature importance of tree model) is ranked in the top 30% during the iteration process. (3) KL divergence time-series sensitivity enhancement screening (threshold > 0.3): ; Among them, P t Let f be the characteristic distribution at time t, P0 be the initial distribution, and f i Let i be the i-th feature.

[0048] Furthermore, according to one embodiment of the present invention, the structure of the model combining bidirectional LSTM and attention mechanism includes: Input layer: Receives multidimensional features, including spatial and temporal features; Bidirectional LSTM layer: It performs bidirectional modeling on the sequence of input multidimensional features, extracting time series features from both the forward and reverse directions; the forward LSTM captures the dependencies from the past to the present, and the reverse LSTM captures the dependencies from the future to the present, thus achieving comprehensive modeling of the time series. Among them, forward LSTM: encodes the historical evolution from the initial time (t=0) to the current time (t=T); ; Reverse LSTM: Backtracks from the current time (t=T) to the initial time (t=0) to capture inverse dependencies; ; Final hidden state: ; d represents the number of hidden units in a single LSTM layer; in this implementation, d = 256. The LSTM gating update mechanism uses a forget gate to determine which historical information has been lost, an input gate to filter out valuable new information and update the state, and an output gate to control the current hidden state. When a segment of an image is of poor quality due to motion artifacts, the forget gate can reduce its weight, while the input gate can enhance features such as tumor volume mutations.

[0049] Attention Mechanism Layer: Based on the bidirectional LSTM output, an attention mechanism is introduced to perform weighted processing on the temporal steps and spatial features between different temporal features; Output layer: The feature vectors processed by the attention mechanism are fused to generate the final prediction result, which is the prediction value of whether or not radiotherapy will be initiated.

[0050] In this embodiment, the adaptive radiotherapy (ART) initiation timing prediction task requires processing two types of key information simultaneously: spatial features and temporal features. Spatial features include changes in the anatomical structure of the target area and organs at risk in each image (such as volume, shape, and texture), while temporal features involve dynamic evolution patterns such as dose accumulation and tumor regression during treatment. Traditional RNNs or single-time-series models have the following drawbacks: unidirectional LSTMs can only capture unidirectional dependencies from history to the present, ignoring potential future trends; without an attention mechanism, the model treats all time steps equally, failing to focus on key treatment nodes (such as sudden tumor regression).

[0051] Based on the analysis of the adaptive initiation timing prediction task, the model not only needs to examine the characteristics of the current moment, but also needs to compare the differences and changes in radiomic features between the current moment and the past state, especially the initial plan, in order to make a reasonable estimate of the probability that adaptive radiotherapy needs to be initiated at the current moment. Therefore, this implementation proposes the aforementioned adaptive spatiotemporal fusion prediction model, which achieves multi-dimensional information fusion through bidirectional LSTM combined with an attention mechanism. This model can simultaneously capture the bidirectional dependencies of time series and use the attention mechanism to weight key time steps and spatial features, thereby improving prediction accuracy and model interpretability.

[0052] Furthermore, according to one embodiment of the present invention, when fusing spatial features, the model first assigns different attention weights to different spatial features (such as target volume, shape, texture, etc.) through an attention mechanism layer, thereby achieving selective fusion of key spatial features. This method not only improves the interpretability of the model but also enhances its ability to capture complex spatiotemporal dependencies.

[0053] Furthermore, according to one embodiment of the present invention, when the model fuses time features, it first calculates the correlation between different time steps through an attention mechanism layer to capture global context information and generate a weighted time state feature representation; then, the time state features are used to perform sequence modeling through an LSTM layer, and a recursive gating mechanism is used to capture long sequences and long-term dependencies, thereby obtaining a richer time feature representation. To address the issue of traditional LSTM treating all time steps "equally," an attention scoring mechanism is introduced into the attention mechanism layer, including: Calculate the energy value: Energy value e t Used to measure the current query vector h T With historical key vector h t The similarity between them is calculated using the following formula: ; Where v is the weight vector, W a and U a It is a learnable parameter matrix, b a It is a bias term; Normalized weights: The attention weight α is obtained by normalizing the energy value using the Softmax function. t This step ensures that the sum of all attention weights is 1, guaranteeing that the model focuses on the historical time step most relevant to the current query; the normalization formula is: ; Generate context vectors: Using attention weight αt For the historical key vector h t Perform a weighted summation to generate a context vector c, which incorporates the information most relevant to the current query: ; In this way, the model can dynamically adjust its focus on different time steps based on the context, thereby improving the accuracy of predictions.

[0054] Input the context vector c into a single neuron layer activated by Sigmoid (i.e., the output layer of the model), and output the ART activation probability: ; When p ART Adaptive radiotherapy is triggered when the value is >0.8.

[0055] Furthermore, according to one embodiment of the present invention, the method of the present invention can take the initial features (radiomics features of the initial image and initial doseomics features) as x0 input, take the fractionated image features (radiomics features of the fractionated image and fractionated doseomics features) at the current time of investigation as xt input, and take the average state avg([x1, x2, ..., xt-1]), the maximum state max([x1, x2, ..., xt-1]), and the minimum state min([x1, x2, ..., xt-1]) of all states {x1, x2, ..., xt-1} between x0 and xt as input, and output the probability of initiating adaptive radiotherapy at the current time after model processing. The state of non-adaptive initiation timing can be used as a negative sample, and the true value of the adaptive initiation probability at that time is set to 0; the state of adaptive initiation timing determined by dose analysis is used as a positive sample, and the true value of its corresponding adaptive initiation probability is set to 1. Based on the collected patient data, the method is trained based on the above-mentioned various feature inputs to provide a reference for adaptive initiation timing decision-making.

[0056] According to the above-described scheme of the present invention, an adaptive radiotherapy initiation decision strategy is designed based on dose analysis and omics information from previous treatments. This scheme allows for the selection of an appropriate timing for initiating adaptive radiotherapy for individualized patients, fully leveraging the advantages of adaptive radiotherapy and further improving treatment outcomes. The present invention can provide each patient with targeted, individualized adaptive initiation timing support, effectively improving the accuracy and quality of treatment for different patients.

[0057] Furthermore, to achieve the above objectives, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the adaptive radiotherapy initiation timing prediction method as described above.

[0058] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the adaptive radiotherapy initiation timing prediction method as described above.

[0059] Those skilled in the art will recognize that the modules and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented using electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0060] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described apparatus and equipment can be referred to the corresponding process in the foregoing method implementation, and will not be repeated here.

[0061] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0062] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the objectives of the embodiments of the present invention, depending on actual needs.

[0063] In addition, the functional modules in the embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0064] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the sending / receiving methods of various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0065] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this application.

[0066] It should be understood that the sequence number of each step in the invention and its embodiments does not absolutely imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

Claims

1. An adaptive radiotherapy initiation timing prediction method, characterized in that, include: The ART index, which quantifies the conditions for initiating radiotherapy, was obtained through image analysis and calculation. Utilize radiomics tools to obtain radiomics characteristics of target areas and key organs at risk in fractionated imaging; The ART index and radiomics features are combined to form a feature set, and the recursive feature elimination combined with KL divergence is used to filter the features in the feature set; The multidimensional features obtained after screening are divided into training set and test set. The model combining bidirectional LSTM and attention mechanism is trained through the training set to form a spatiotemporal fusion prediction model. The test set is input into the spatiotemporal fusion prediction model, and the predicted timing of radiotherapy initiation is output.

2. The adaptive radiotherapy initiation timing prediction method according to claim 1, characterized in that, The ART metric includes an initial ART metric and a phased ART metric, both of which include: Tumor volume shrinkage rate; Planned target area dose deviation; The rate of overdose on organs at risk; Biological effects TCP and NTCP.

3. The adaptive radiotherapy initiation timing prediction method according to claim 1, characterized in that, The radiomics features include: Patient's clinical diagnostic characteristics; Radiomic features of the initial images; Initial dose-omics characteristics; Radiomics features of fractional images; fractional dose omics characteristics; Cumulative dose-omics characteristics.

4. The adaptive radiotherapy initiation timing prediction method according to claim 1, characterized in that, The method of using recursive feature elimination combined with KL divergence to filter features from the feature set includes: (1) Calculate the delta eigenvalue and retain the eigenvalues ​​with time series fluctuations >30%; (2) Recursive feature elimination is used for initial screening, retaining the top 30% of features; (3) KL divergence time-series sensitivity enhancement screening: ； Among them, P t Let f be the characteristic distribution at time t, P0 be the initial distribution, and f i Let i be the i-th feature.

5. The adaptive radiotherapy initiation timing prediction method according to claim 1, characterized in that, The structure of the model combining bidirectional LSTM and attention mechanism includes: Input layer: Receives multidimensional features, including spatial and temporal features; Bidirectional LSTM layer: It performs bidirectional modeling on the sequence of input multidimensional features, extracting time series features from both the forward and reverse directions; the forward LSTM captures the dependencies from the past to the present, and the reverse LSTM captures the dependencies from the future to the present, thus achieving comprehensive modeling of the time series. Attention Mechanism Layer: Based on the bidirectional LSTM output, an attention mechanism is introduced to perform weighted processing on the temporal steps and spatial features between different temporal features; Output layer: The feature vectors processed by the attention mechanism are fused to generate the final prediction result, which is the prediction value of whether or not radiotherapy will be initiated.

6. The adaptive radiotherapy initiation timing prediction method according to claim 5, characterized in that, When fusing spatial features, the model first assigns different attention weights to different spatial features through an attention mechanism layer, thereby achieving selective fusion of key spatial features.

7. The adaptive radiotherapy initiation timing prediction method according to claim 6, characterized in that, When fusing temporal features, the model first calculates the correlation between different time steps through an attention mechanism layer to capture global context information and generate a weighted temporal state feature representation; then, the temporal state features are used to perform sequence modeling through an LSTM layer, and a recursive gating mechanism is used to capture long sequences and long-term dependencies. The attention mechanism layer introduces an attention scoring mechanism, including: Calculate the energy value: Energy value e t Used to measure the current query vector h T With historical key vector h t The similarity between them is calculated using the following formula: ； Where v is the weight vector, W a and U a It is a learnable parameter matrix, b a It is a bias term; Normalized weights: The attention weight α is obtained by normalizing the energy value using the Softmax function. t To ensure that the sum of all attention weights is 1, the normalization formula is: ； Generate context vectors: Using attention weight α t For the historical key vector h t A weighted sum is performed to generate a context vector c, which incorporates the information most relevant to the current query. ； Input the context vector c into a single neuron layer activated by Sigmoid, and output the ART activation probability: ； When p ART Adaptive radiotherapy is triggered when the value is >0.

8.

8. An adaptive radiotherapy initiation timing prediction system, characterized in that, include: The ART index acquisition module calculates ART indexes that quantify the conditions for initiating radiotherapy through image analysis. The radiomics feature acquisition module uses radiomics tools to acquire radiomics features of target areas and key organs at risk in fractional imaging. The feature selection module combines ART indicators and radiomics features into a feature set, and uses recursive feature elimination combined with KL divergence to select features from the feature set. The spatiotemporal fusion prediction model construction module divides the selected multidimensional features into training and testing sets. The training set is used to train the model that combines bidirectional LSTM with attention mechanism to form a spatiotemporal fusion prediction model. The test set is input into the spatiotemporal fusion prediction model, and the predicted timing of radiotherapy initiation is output.

9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the adaptive radiotherapy initiation timing prediction method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the adaptive radiotherapy initiation timing prediction method as described in any one of claims 1-7.

Images