Task-adaptive heterogeneous multi-modal fusion based lung cancer radiotherapy outcome parameter generation method
By employing a task-adaptive heterogeneous multimodal fusion method, feature subsets and generation parameters are constructed for different outcome types of lung cancer radiotherapy data. This solves the problems of redundant interference and inconsistent processing rules under multi-source heterogeneous data, and achieves more accurate radiotherapy outcome prediction.
Patent Information
- Application Number
- CN202611598827.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-15
- Publication Date
- 2026-11-13
AI Technical Summary
Given the multi-source heterogeneous data generated by lung cancer radiotherapy, existing technologies struggle to generate feature representations and parameter generation processes adapted to their respective tasks due to the differences in data structure between radiotherapy response classification outcomes and progression-free survival censoring outcomes. This leads to problems such as redundant interference and inconsistent processing rules.
A task-adaptive heterogeneous multimodal fusion method for generating radiotherapy outcome parameters in lung cancer is proposed. By acquiring clinical parameters, radiotherapy dose parameters, CT radiomics parameters, and TP53I3 molecular pathology scores of the same patient, four-modal data are matched according to patient identifiers to construct a unified heterogeneous multimodal feature pool. Task-adaptive routing is then performed according to the outcome parameter type, calling either the radiotherapy response classification branch or the progression-free survival analysis branch. Sparse screening with regularization constraints and survival model ranking are used to form feature subsets, generating radiotherapy response probability and progression-free survival risk parameters.
It reduces redundant interference from high-dimensional feature sets on outcomes with different statistical properties, maintains consistency in data processing during training, validation, and independent application, and improves the model's prediction accuracy and stability.
Smart Images

Figure CN123095261A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing and medical image analysis technology for lung cancer radiotherapy, specifically to a method for generating lung cancer radiotherapy outcome parameters using task-adaptive heterogeneous multimodal fusion. Background Technology
[0002] Lung cancer is one of the most common malignant tumors, with non-small cell lung cancer accounting for the majority of lung cancer cases. For patients who cannot undergo surgery or whose primary local treatment is radiotherapy, radiotherapy is an important treatment option. Due to differences in individual patient conditions, tumor biological characteristics, and radiotherapy-related factors, even with similar radiotherapy doses and treatment regimens, different patients may still experience varying treatment responses, disease progression time, and subsequent survival outcomes.
[0003] With the development of medical informatics and radiotherapy technology, lung cancer radiotherapy generates data from various sources. Clinical parameters reflect the patient's general condition, physical condition, lung function, tumor stage, and serological information. Radiotherapy dose parameters reflect target dose coverage, dose homogeneity, dose gradient, treatment plan complexity, and dose received by normal lung tissue. CT images can also be used to generate quantitative features such as shape, first-order statistics, and texture through radiomics methods. Molecular pathology results can provide supplementary information from a tumor biology perspective. Therefore, joint analysis of data from these different sources can provide a multidimensional information foundation for generating radiotherapy outcome parameters.
[0004] Chinese invention patent application CN115564735A (publication date January 3, 2023) discloses a method for predicting nasopharyngeal necrosis after radiotherapy for recurrent nasopharyngeal carcinoma. This method acquires multimodal MRI data, extracts radiomics features from the tumor region, preprocesses radiotherapy dose data, and extracts dose features using a 3D CNN. The radiomics features and dose features are then fused to generate the probability of radiotherapy-induced necrosis. This method demonstrates that medical imaging features and radiotherapy dose features can be used together for radiotherapy-related outcome analysis, but its processing flow is established around the single outcome of the probability of radiotherapy-induced necrosis.
[0005] Chinese invention patent application CN118314106A (publication date July 9, 2024) discloses a prognostic prediction method for non-small cell lung cancer based on multimodal fusion. This method acquires pre-treatment CT images and three-dimensional dose distribution images from radiotherapy, extracts radiomics and dosimics features from regions of interest, forms multi-omics prediction labels and deep survival labels, and combines these labels with prognostic-related clinical features for progression-free survival (PFS) prediction. This method combines CT, radiotherapy dose, and clinical information, but its multi-omics prediction, deep survival prediction, and final fusion are all built around the PFS outcome.
[0006] In some multimodal prediction processes, data from different sources often have significantly different dimensions. Radiomics data can contain a large number of high-dimensional variables, while clinical parameters and molecular pathology data have relatively low dimensions. If a fixed set of features is directly used for fusion, the high-dimensional modalities are prone to introducing a lot of redundant information and affecting the information contribution of other low-dimensional modalities.
[0007] Furthermore, the data structures for different radiotherapy outcomes are not the same. Short-term radiotherapy response is usually represented as a categorical outcome, while progression-free survival involves not only whether progression occurred but also the timing of the progression event and censored data. If the same feature set and processing structure are used for outcomes with different statistical properties, it is difficult to retain information that is highly correlated with each outcome, and it is also not conducive to maintaining consistency in data processing during model training, validation, and independent application.
[0008] Therefore, under the condition of multi-source heterogeneous data generated by lung cancer radiotherapy, it is necessary to develop information representation and parameter generation processes that are adapted to the data structure for radiotherapy outcomes with different statistical properties, so as to reduce interference from high-dimensional redundant information. In the process of model training, validation and independent application, it is also necessary to avoid validation data or external data from participating in the formation of training parameters, so that each stage adopts data processing rules determined by training data. Summary of the Invention
[0009] The technical problem this invention aims to solve is: under the condition of multi-source heterogeneous data generated by lung cancer radiotherapy, to form feature representations and parameter generation processes adapted to their respective tasks based on the differences in data structure between radiotherapy response classification outcomes and progression-free survival censoring outcomes, thereby reducing redundant interference caused by using the same high-dimensional feature set; and further solving the problem that inconsistencies in processing rules may occur during the training, validation, and independent application stages due to re-estimation of data processing parameters or re-selection of features.
[0010] To address the aforementioned technical problems, this invention proposes a task-adaptive heterogeneous multimodal fusion method for generating radiotherapy outcome parameters for lung cancer, comprising: acquiring clinical parameters, radiotherapy dose parameters, CT radiomics parameters with the clinical target area as the volume of interest, and TP53I3 molecular pathology score of the same patient; Match four-modal data according to patient identifiers, complete missing value and data scaling processing, and construct a unified heterogeneous multimodal feature pool under the same case index; Adaptive routing is performed based on the outcome parameter type, calling either the radiotherapy response classification branch or the progression-free survival analysis branch respectively. Radiation therapy response branches utilize response category labels, which are based on response type labels. The sparse screening of regularization constraints preserves features with non-zero coefficients, forming a subset of features for the first task; class balancing is performed within the training subset, integrating the prediction results of at least two classification models to generate radiotherapy response probability parameters. The no-progression survival branch sorts and filters candidate features based on survival time and progress event status using a Cox survival model with regularization constraints to form a subset of features for the second task; it then integrates the risk outputs of at least two survival analysis models to generate no-progression survival risk parameters. After training, missing values, scale parameters, task feature indexes, and model parameters are fixed; these parameters are used for new patients without re-selecting features, and external data is not used for model refitting. Output at least one of a radiotherapy response probability parameter and a progression-free survival risk parameter; wherein the two task feature subsets are formed based on a classification objective and a censored survival objective, respectively, and differ in at least one of modality composition, specific features, or feature weights.
[0011] Preferably, missing value processing and feature encoding are performed on each modality data, with missing value processing parameters determined by the training data; the mean of the continuous variable is determined based on the training data. and standard deviation ,according to Scale processing is performed, among which, To process the eigenvalues beforehand, These are the processed eigenvalues. The mean of the training data for the aforementioned feature. The standard deviation of the training data for the aforementioned feature is used; the corresponding processing parameters are fixed and applied to the data to be processed outside of the training data.
[0012] Preferably, the clinical parameter data includes at least one of the following parameters: age, sex, smoking status, pathological type, ECOG score, percentage of predicted forced expiratory volume in one second, percentage of predicted carbon monoxide diffusion capacity in the lungs, TNM stage, NSE, CYFRA21-1, CEA, SCC, and CA153.
[0013] Preferably, the radiotherapy dose parameter data includes at least one of the following parameters: fractionation parameter, beam parameter, treatment plan complexity parameter, minimum dose to clinical target area, maximum dose to clinical target area, average dose to clinical target area, conformity index, homogeneity index, dose gradient index, average lung dose, and lung tissue dose. and lung tissue The and Each characterizes lung tissue receiving at least and Volumetric parameters of irradiation.
[0014] Preferably, the acquisition of the radiomics data includes: spatial resampling and fixed gray-width discretization of the CT image using the clinical target area as the volume of interest, and extracting shape features, first-order statistical features, gray-level co-occurrence matrix features, gray-level dependency matrix features, gray-level run-length matrix features, gray-level region size matrix features, and adjacent gray-level difference matrix features from the original CT image and the transformed CT image; wherein, the voxel size of the spatial resampling is... Fixed grayscale width is .
[0015] Preferably, the TP53I3 molecular pathological scoring data is obtained by digitally encoding the four-level expression results of TP53I3 recorded in the pathological test results. The four-level expression results are no expression, low expression, moderate expression and high expression, respectively, and are encoded as 0, 1, 2 and 3.
[0016] Preferably, stratified five-fold cross-validation is used in the radiotherapy response task. In each cross-validation cycle, missing value handling parameters and scaling parameters are determined within the training subset, using parameters with... The sparse feature selection method with regularization constraints retains the features corresponding to non-zero coefficients to form the first task feature subset, and performs class balancing and classification model training by synthesizing minority class oversampling method; the corresponding validation subset does not participate in parameter determination, feature selection and class balancing.
[0017] Preferably, the at least two classification models are selected from logistic regression, random forest, LightGBM, and XGBoost models; the out-of-division prediction probabilities output by each basic classification model for training samples that did not participate in the current fitting are obtained through cross-validation; the out-of-division prediction probabilities of the same training sample are combined into secondary inputs according to the model order; and the secondary inputs are processed by the Stacking secondary fusion model to generate radiotherapy response probability parameters.
[0018] Preferably, in the no-progression survival task, a Cox survival model with regularization constraints is used. Features are ranked based on their association with no-progression survival time and progress event states, and the top-ranked features are selected. The candidate features form a subset of the second task features; among them... This indicates the number of features selected, and .
[0019] Preferably, the second task feature subset is input into the random survival forest model and the gradient boosting survival analysis model, respectively, to obtain risk outputs. and ,according to Weighted fusion is performed, where, For progression-free survival risk parameters, For the fusion weights of the random survival forest model, and .
[0020] Compared with the prior art, the present invention has the following beneficial effects: Compared to existing technologies that establish fixed data processing flows around a single outcome, this application first performs case-level matching on clinical parameter data, radiotherapy dose parameter data, CT radiomics data, and TP53I3 molecular pathology score data of the same target subject to form a unified heterogeneous multimodal feature pool. Then, based on the outcome parameter type, these pools are respectively assigned to the radiotherapy response task branch and the progression-free survival task branch. Since the two tasks form task feature subsets based on classification and censored survival objectives respectively, redundant interference caused by mechanically applying the same high-dimensional feature set to outcomes with different statistical properties can be reduced. The radiotherapy response task integrates the prediction results of at least two classification models, and the progression-free survival task fuses the risk outputs of at least two survival analysis models. In a further embodiment, the data processing parameters, task feature indexes, and model parameters formed during the training phase remain fixed during the independent application phase, thereby maintaining data processing consistency across training, validation, and independent application processes, and reducing the risk of information leakage caused by external data participating in the formation of training parameters. Attached Figure Description
[0021] Figure 1 The diagram above illustrates the overall structure of the lung cancer radiotherapy outcome parameter generation method based on task-adaptive heterogeneous multimodal fusion provided in this application embodiment. It sequentially represents the acquisition of four types of modal data, case-level matching and feature pool construction, task-adaptive routing, generation of radiotherapy response probability parameters, generation of progression-free survival risk parameters, and output of outcome parameters. Detailed Implementation
[0022] The technical solution of this application will be described below through specific embodiments. It should be understood that the following embodiments are used to illustrate the technical solution of this application, and not to limit the scope of protection of this application. Without changing the technical concept of this application, those skilled in the art can implement corresponding data acquisition, storage, transmission and program calling according to actual data conditions and computing environment.
[0023] Example 1 This embodiment provides a task-adaptive heterogeneous multimodal fusion method for generating radiotherapy outcome parameters for lung cancer. This method can be executed by a computer system with data processing capabilities. The computer system's processor calls a pre-stored program to process clinical parameter data, radiotherapy dose parameter data, radiomics data, and TP53I3 molecular pathology score data for the same target object, and executes different data processing branches according to the type of radiotherapy outcome parameter to be generated.
[0024] like Figure 1 As shown, this embodiment first establishes a unified case-level heterogeneous multimodal data foundation, and then performs adaptive task routing based on the data structure of the outcome parameters to be generated. When the target is a radiotherapy response parameter, feature selection and classification model integration for categorical outcomes are performed. When the target is a progression-free survival risk parameter, feature selection and survival risk fusion for censored survival data are performed. Both tasks use the same case-level multimodal data foundation as input, but do not require the use of the same final feature set. Figure 1 The first flowchart corresponds to acquiring clinical parameters, radiotherapy dose parameters, clinical target area radiomics data, and TP53I3 score data; the second flowchart corresponds to case-level matching and feature pool construction; the third flowchart corresponds to calling the processing branch according to the outcome type; the fourth and fifth flowcharts correspond to the classification prediction branch and the censored survival risk branch, respectively; and the sixth flowchart corresponds to outputting the corresponding outcome parameters.
[0025] This embodiment targets non-small cell lung cancer patients undergoing radical radiotherapy. In one implementation of model development, a model development cohort of 190 patients can be used. The radiotherapy response task and progression-free survival task use the same case-level data base, ensuring that modal data for the same target subject remain consistent across different tasks.
[0026] First, clinical parameter data is obtained. Clinical parameters can be obtained from hospital clinical information systems, electronic medical record systems, and pulmonary function test results. In one implementation, the clinical parameters comprise 16 dimensions, covering basic patient information, physical status, pulmonary function, tumor stage, and tumor markers. Specifically, at least one of the following parameters can be selected: age, sex, smoking status, pathological type, ECOG score, percentage of predicted forced expiratory volume in one second (FEV1), percentage of predicted carbon monoxide diffusing capacity (COD), TNM stage, NSE, CYFRA21-1, CEA, SCC, and CA153.
[0027] The percentage of predicted forced expiratory volume in one second (FEV1) is used to characterize airway ventilation reserve; the percentage of predicted carbon monoxide diffusing capacity (COD) in the lungs is used to characterize the gas diffusion capacity between alveoli and capillaries; and the ECOG score is used to characterize the overall functional status of the target subject. These clinical data, after being structured, are input as clinical parameter features into subsequent data processing workflows.
[0028] Further radiotherapy dose parameter data is obtained. This data can come from the radiotherapy plan or relevant dose data exported from the radiotherapy planning system. In one embodiment, the dose-volume histogram (DVH) and planning-related parameters comprise 26 dimensions, and the parameters can be selected from at least one of the following: fractionation parameters, beam parameters, treatment plan complexity parameters, minimum CTV dose, maximum CTV dose, average CTV dose, conformity index, homogeneity index, dose gradient index, average dose for both lungs or one lung, and lung tissue... and lung tissue .in, and These respectively indicate that the corresponding lung tissue received no less than and Volumetric parameters of irradiation.
[0029] Radiotherapy dose parameters not only reflect the dose level received by the target area, but also characterize dose coverage, spatial conformity, dose distribution uniformity, dose gradient, dose received by normal lung tissue, and the geometric complexity of the treatment plan. For example, the RTOG conformity index is used to characterize the spatial coverage relationship of the prescribed dose to the target area, and the aperture irregularity metric is used to characterize the geometric complexity of the irradiation field formed by a multi-leaf collimator. The obtained dose and planning parameters are presented as numerical characteristics to form radiotherapy dose parameter data.
[0030] Further acquisition of treatment-related CT images and their corresponding CTVs is required. The CTVs can be derived from target delineation results confirmed by clinicians within a radiotherapy planning system. Spatial correspondence between the CT images and CTVs is established, and the CTVs are used as the volume of interest for radiomics feature extraction.
[0031] Before radiomics feature extraction, CT images undergo uniform spatial and grayscale processing. In a preferred embodiment, the CT images are isotropically resampled, and the resampled voxel size is [missing information]. The image is then discretized using a fixed grayscale width. .
[0032] After preprocessing, quantitative radiomics features are extracted from the original CT images and the transformed images. The extracted features may include shape features, first-order statistical features, gray-level co-occurrence matrix features, gray-level dependency matrix features, gray-level run-length matrix features, gray-level region size matrix features, and adjacent gray-level difference matrix features. In one experimental implementation, 1409-dimensional radiomics raw features are formed from the CTV region. These 1409-dimensional features serve as candidate radiomics data before task-specific screening and do not imply that all radiomics features must be used in the subsequent two tasks.
[0033] Further, TP53I3 pathological detection results of the same target object are obtained, and the TP53I3 four-level expression results recorded in the pathological detection results are read and converted into a digital molecular pathology score that can be processed by a computer. In one embodiment, the four-level expression results are, in order, no expression, low expression, moderate expression, and high expression, and are encoded as 0, 1, 2, and 3, respectively. Through this processing, the TP53I3 expression results in pathological tissues are converted into numerical features that can be processed together with clinical parameters, dose parameters, and radiomics parameters.
[0034] After obtaining the above four types of modal data, case-level matching is performed based on the unique identifier of the target object. Clinical parameter data is denoted as... Radiotherapy dose parameter data are recorded as Radiomics data are denoted as TP53I3 molecular pathology score data is recorded as After case-level mapping, a unified heterogeneous multimodal feature pool is formed. It is represented as:
[0035] In one implementation, the original total dimension of the unified heterogeneous multimodal feature pool is 1452. This includes 16 dimensions of clinical parameters, 26 dimensions of radiotherapy dose parameters, 1409 dimensions of radiomics parameters, and 1 dimension of TP53I3 score. Data from different dimensions are used to form feature records corresponding to the same target object under a unified case index. This represents the complete multimodal candidate feature vector corresponding to the same case; , , , These represent the feature components corresponding to clinical parameters, radiotherapy dose parameters, clinical target area radiomics parameters, and TP53I3 molecular pathology score parameters, respectively. Square brackets indicate the combination of features according to case identification and a predetermined order.
[0036] For a small number of missing values in the data, during the model training phase, only the corresponding training data is used to determine the parameters needed for missing value handling, and these parameters are then applied to the corresponding validation data. Similarly, for continuous variables requiring scaling, only the training data is used to calculate the mean. and standard deviation Perform the Z-score transformation according to the following formula:
[0037] in, This represents the feature values before scaling. This represents the eigenvalues after scaling. This represents the mean of the corresponding feature obtained from the training data. This represents the standard deviation of the corresponding feature obtained from the training data. It is determined during the training phase. and It remains constant during subsequent independent data processing and is not recalculated based on newly input external data.
[0038] After constructing a unified heterogeneous multimodal feature pool, adaptive task routing is executed based on the type of lung cancer radiotherapy outcome parameters to be generated. Task identifiers are then set. This is used to indicate the current probability parameter of generating a radiotherapy response or the risk parameter of progression-free survival. When When instructing the generation of radiotherapy response probability parameters, the unified heterogeneous multimodal feature pool is routed to the first task branch. When generating no-progression survival risk parameters, the unified heterogeneous multimodal feature pool is routed to the second task branch. When two outcome parameters need to be obtained simultaneously, the unified heterogeneous multimodal feature pools corresponding to the same target object are input into the two task branches respectively.
[0039] Task adaptive routing does not merely change the final output node. The two task branches have supervision objectives and feature selection processes corresponding to their respective outcome data structures. The first task branch forms a first task feature subset based on the classification relationship between responses and non-responses. The second task branch forms a subset of features based on the no-progression survival time and the status of censorship events. Therefore, the first task feature subset and the second task feature subset do not need to be the same. They can differ in at least one of the following: the modality used, the specific features retained, or the corresponding feature weights.
[0040] For the radiotherapy response task, training labels are first established based on pre-determined clinical radiotherapy response evaluation results. Valid responses are coded into one class, and no valid responses into another. Hierarchical five-fold cross-validation is used during model development to ensure that the proportion of samples from different classes remains approximately consistent within each fold.
[0041] For the 1 cross-validation cycle, of which The index of the cross-validation period indicates the number of the training subsets that divide the data into non-overlapping training subsets. and verification subset Missing value handling parameters, scaling parameters, first task feature selection, and model parameter determination related to model training are all included in... Internally, this is performed, and class balancing is achieved within the training subset using Synthetic Minority Oversampling (SMOTE). The data transformation rules formed from the training subset are then applied. The validation subset does not participate in the formation of training parameters.
[0042] To classify radiotherapy responses, a system is established within the training subset that has the following characteristics: A sparse feature selection model with regularization constraints; in one embodiment, the sparse feature selection model adopts the LASSO model. Its optimization relationship can be expressed as:
[0043] in, This represents the feature coefficient vector obtained through optimization. This represents the feature coefficient vector to be optimized. This represents the candidate feature matrix of the classification model. This represents the corresponding linear prediction. Represents the classification loss function. This indicates the corresponding radiotherapy response category label. express Regularization strength, Indicates the first The coefficients corresponding to each candidate feature Iterate through all candidate features, and the summation term represents the sum of the absolute values of the coefficients. Regularization constraints are applied to compress the coefficients of candidate features that contribute little to the current classification task or have significant redundancy. Features with non-zero coefficients are selected as the first subset of task features.
[0044] In one experimental implementation, 17 non-zero features were obtained through task-specific screening as described above. These 17 features cover clinical parameters, DVH parameters, radiomics parameters, and the TP53I3 molecular pathology score. Representative features with high model gain include the percentage of predicted forced expiratory volume in one second (FEV1), the percentage of predicted carbon monoxide diffusion in the lungs, the CTV_RTOG Conformity Index, the original_glszm_SizeZoneNonUniformityNormalized index, the irradiation field irregularity index, the TP53I3 score, the original_glcm_Autocorrelation, the ECOG score, the mean lung dose, and NSE. These 17 features are the result of task screening of the experimental data in this embodiment and are not a pre-fixed input set for a unified heterogeneous multimodal feature pool.
[0045] After obtaining the first subset of task features, these features are input into multiple base classification models. In one implementation, the base classification models include logistic regression, random forest, LightGBM, and XGBoost models. Alternatively, at least two of these models can be selected for ensemble integration.
[0046] For each basic classification model, cross-validation is used to generate the out-of-bounds prediction probability of the training samples. For the th 1 target object, of which The sample index represents the out-of-place prediction probabilities corresponding to each basic classification model, which can be used as secondary inputs. This refers to the vector composed of the out-of-division predicted probabilities of each base classifier for the same sample. When using the four base classification models—logistic regression, random forest, LightGBM, and XGBoost— Represented as:
[0047] in, , , and These represent the four basic classification models—Logistic Regression, Random Forest, LightGBM, and XGBoost—for the first... The out-of-bounds predicted probability of each target object.
[0048] Secondary input Input the two-level fusion model obtained from training ,in, This represents a second-level fusion function that takes the out-of-fold prediction probabilities of the base classifiers as input; it fuses the prediction information of multiple base classification models through stacking ensemble to obtain the second-level fusion function. Radiotherapy response probability parameters for a target individual :
[0049] Therefore, the radiotherapy response probability parameter is not directly obtained from the entire high-dimensional unified feature pool through a single fixed model. Instead, driven by the classification task objective, a first task feature subset is first formed, and then fused using prediction information from multiple basic classification models. This process allows clinical, dosimetry, radiomics, and molecular pathology modalities to participate in parameter generation based on their correlation with the radiotherapy response classification results. Indicates the first Predicted probability of radiotherapy response for a target individual This represents the combination of predicted probabilities from the base classification model for the target object. This indicates a trained two-level fusion model.
[0050] In one experimental dataset, the trimodal control group consisted of clinical parameters, DVH parameters, and radiomics parameters, while the quadruple control group added the TP53I3 molecular pathology score to the trimodal control group. When performing five-fold cross-validation comparisons of different modalities in the radiotherapy response task, the area under the receiver operating characteristic (ROC) curve (AUC) for the TP53I3 monomodal, clinical parameter monomodal, trimodal control group, clinical parameter and TP53I3 combination, clinical parameter and DVH parameters and TP53I3 combination, clinical parameter and radiomics parameters and TP53I3 combination, and quadruple control group were approximately [missing data]. , , , , , and The clinical parameters showed that the combination of TP53I3 and clinical parameters improved the AUC by approximately [missing information] compared to the trimodal control combination. In another set of model iteration comparisons, the AUCs of the historically best three-modal scheme, the TP53I3 single-modal scheme, the combination of clinical parameters and TP53I3, the combination of clinical parameters, DVH parameters, and TP53I3, the hyperparameter-optimized LightGBM model, and the final four-modal Stacking model are approximately, respectively, approximately , , , , and The final five-fold cross-validation AUC of the Stacking model based on four modalities and 17 features is approximately [value missing]. The mean AUC obtained by bootstrap resampling is approximately , The range is approximately Sensitivity is approximately The specificity is approximately In one experimental result of 500 permutation tests, unilateral The value is approximately .
[0051] For no-progression survival tasks, the no-progression survival time is recorded for each target object. and event status When a preset progress event occurs in the target object during the observation period, let... If the preset progress event does not occur before the end of the observation, then... This results in censored survival data that includes uniform heterogeneous multimodal features, progress-free survival time, and event states. :
[0052] in, This represents a censored surviving dataset. Indicates the first Multimodal features of a target object This indicates the progressless survival time of the target object. Indicates the progress status of the event. Index the target object; This indicates that the aforementioned progress event occurred during the observation period. This indicates that the event has not occurred by the end of the observation period. Compared to the binary label for radiotherapy response, progression-free survival data includes both whether the event occurred and when it occurred. Therefore, the second task branch does not directly use the classification objective function of the first task branch, but instead uses an objective function capable of handling censored survival data to evaluate candidate features.
[0053] Specifically, a Cox survival model with regularization constraints is used to evaluate the relationship between unified heterogeneous multimodal features and censored survival outcomes; in one implementation, the Cox survival model with regularization constraints is the CoxNet model. The corresponding risk function can be expressed as:
[0054] in, Represents a given multimodal feature In time The risk function, Indicates the observation time. Represents the baseline risk function. This represents the survival association parameter vector corresponding to the multimodal candidate features. This represents the transpose of the vector. Represents an exponential function; to This represents candidate features ranked by their degree of survival association, where This represents the total number of candidate features. Candidate features are ranked according to regularization path, non-zero coefficient, or feature importance, forming... Then select the one that is ranked first. These features form a subset of the second task features. .
[0055] In a preferred embodiment, ,in This indicates the number of candidate features retained in the ranking. In the experimental data, the top 50 features in the second task feature subset were mainly CTV radiomics features, while also including DVH parameters and clinical parameters. This feature composition is different from the 17 features selected in the radiotherapy response task, thus enabling the same case-level data base to form feature representations applicable to both categorical and censored survival outcomes.
[0056] Obtain the second task feature subset Then, the data were input into a Random Survival Forest (RSF) model and a Gradient Boosting Survival Analysis (GBSA) model, respectively. The RSF model constructs multiple survival trees based on the censored survival data, obtains risk relationships in different samples and feature subspaces, and outputs risk parameters. The GBSA model progressively builds a sequence of weak learners based on the optimization direction of the survival loss and outputs risk parameters. .
[0057] The risk parameters output by the two survival analysis models are weighted and fused to obtain the progression-free survival risk parameters. :
[0058] in, The fusion parameter represents the risk of no progress in survival. This represents the risk output of the random survival forest model. This indicates the risk output of the gradient boosting survival analysis model. This represents the fusion weights determined during the training phase of the random survival forest model, and satisfies... In a preferred embodiment, At this point, the fusion weights corresponding to the random survival forest model are: The fusion weights corresponding to the gradient boosting survival analysis model are: .
[0059] In one experimental dataset, when comparing model iteration results for a progression-free survival task, the consistency index (C-index) of the clinical feature Cox proportional hazards model baseline, the clinical feature RSF model baseline, the four-modal CoxNet model, the hyperparameter-optimized three-modal GBSA model, the hyperparameter-optimized four-modal GBSA model, the hyperparameter-optimized four-modal RSF model, and the final four-modal weighted ensemble model were approximately... , , , , , and The C-index of the five-fold cross-validation of the RSF and GBSA weighted fusion model formed in the above manner is approximately [value missing]. The trimodal GBSA model used clinical parameters, DVH parameters, and radiomics parameters, while the quadmodal model added the TP53I3 molecular pathology score. After 1000 Bootstrap resampling iterations, the mean C-index was approximately [value missing]. , The range is approximately ; obtained after 150 permutation tests The value is approximately These results demonstrate that the second task branch can generate individual risk ranking parameters based on censored survival data.
[0060] In this embodiment, the first and second task feature subsets are formed by different supervision objectives. The first task feature subset is sparsely filtered based on response classification error, while the second task feature subset is filtered considering both progression-free survival time and event status. Therefore, even though both come from the same unified heterogeneous multimodal feature pool, they can form different modal compositions, different specific feature compositions, or different feature weights, thereby avoiding the mechanical application of identical high-dimensional feature sets to two radiotherapy outcome tasks with different statistical properties.
[0061] After model training is complete, the missing value handling parameters, scaling parameters, task feature indexes, classification model parameters, stacking fusion parameters, survival analysis model parameters, and survival risk fusion parameters generated during the training phase are solidified into model configuration data. This model configuration data is stored by the computer system and used in subsequent parameter generation processes.
[0062] For new target objects outside the training data, clinical parameter data, radiotherapy dose parameter data, CTV radiomics data, and TP53I3 molecular pathology score data are first acquired according to the data definition consistent with the model training phase, and case-level matching is completed according to the target object identifier. Subsequently, the data transformation is performed by directly calling the fixed missing value handling parameters, scaling parameters, and task feature index, without using new target objects or independent external data to re-select features, or refitting the classification model, survival analysis model, or fusion model.
[0063] When task identifier When instructing a radiotherapy response task, the first task features corresponding to the new target object are extracted according to the fixed first task feature index, and the already trained basic classification model and Stacking fusion model are invoked to obtain the radiotherapy response probability parameters. When the task identifier... When instructing a progression-free survival task, second-task features are extracted according to the fixed second-task feature index, and the already trained RSF model, GBSA model, and corresponding fusion weights are invoked to obtain progression-free survival risk parameters. When both parameters are needed simultaneously, the above two branches are executed separately for the same case-level feature pool.
[0064] In one independent validation implementation, a separate external cohort was formed using 90 target subjects who did not participate in feature selection and parameter optimization during model training. External data were case-level aligned according to patient identification for clinical parameters, DVH parameters, CTV radiomics parameters, and TP53I3 scores, and the data processing parameters, task feature sets, and model configurations already fixed during training were fully reused. The AUC of the ROC curve for the radiotherapy response model in this independent external cohort was approximately [value missing]. The accuracy rate is approximately , The fraction is approximately Sensitivity is approximately The specificity is approximately The AUC of the ROC curves for internal validation and the three-modal baseline are both approximately [value missing]. In the precision-recall evaluation, the mean precision (AP) for the trimodal baseline, the quadruple TP53I3 internal validation, and the external validation were approximately [missing information]. , and The corresponding baseline positivity rate was The C-index of the no-progression survival risk fusion model in this independent external cohort is approximately [value missing]. Ninety subjects were divided into a high-risk group and a low-risk group, with 45 subjects in each group, according to predetermined risk parameters. The Kaplan-Meier survival curves of the two groups separated, with the overall PFS probability of the high-risk group being lower than that of the low-risk group; the log-rank test... The value is approximately .
[0065] Through the above implementation process, based on the same clinical, dose-related, radiomics, and molecular pathology data, the radiotherapy response task and the progression-free survival task are formed into different task feature subsets according to their respective data structures, and parameter generation processes matching the corresponding tasks are invoked. The data transformation rules, task feature indices, and model parameters obtained during the training phase remain fixed during independent deployment, thereby ensuring that the training data and subsequent target objects adopt a consistent data processing and forward computation process.
[0066] Example 2 This embodiment, based on Embodiment 1, further explains the construction process of a unified heterogeneous multimodal feature pool. This embodiment focuses on the case-level organization and data consistency methods for data from different sources and dimensions before they enter the task adaptive processing stage.
[0067] For the same target group, separate clinical parameter data were established. Radiotherapy dose parameter data radiomics data and TP53I3 molecular pathology scoring data Different modalities of data are mapped using the same target object identifier to ensure that different modalities of data in the same case record originate from the same target object. , , and The meaning is the same as in Example 1.
[0068] Clinical parameters are used to describe the target subject's basic condition, lung function status, tumor characteristics, and serological characteristics. In one implementation, the clinical parameters comprise 16 dimensions, and at least one can be selected from the following parameters: age, sex, smoking status, pathological type, ECOG score, percentage of predicted forced expiratory volume in one second (FEV1), percentage of predicted carbon monoxide diffusing capacity (COD), TNM stage, NSE, CYFRA21-1, CEA, SCC, and CA153. For categorical clinical variables, they are converted into preset codes that can be processed by the computer model. For continuous clinical variables, the data scaling parameters determined by the training data are used for conversion in subsequent data scaling processing.
[0069] Radiotherapy dose parameters are obtained from the radiotherapy plan or corresponding dose data. In one embodiment, the DVH and treatment plan-related parameters comprise 26 dimensions. These parameters can be selected from at least one of the following: fractionation parameters, beam parameters, treatment plan complexity parameters, minimum CTV dose, maximum CTV dose, average CTV dose, RTOG conformity index, homogeneity index, dose gradient index, average dose to both lungs or one lung, and lung tissue... and lung tissue .
[0070] The minimum CTV dose, maximum CTV dose, and average CTV dose are used to numerically characterize the dose received in the target area. The conformity index is used to numerically characterize the spatial coverage of the prescribed dose. The homogeneity index is used to numerically characterize the dose uniformity within the target area. The dose gradient index is used to numerically characterize the dose decrease characteristics around the target area. Lung mean dose and lung tissue... and lung tissue Used to generate dose information for normal lung tissue. Treatment plan complexity parameters are used to generate numerical information related to the geometric complexity or execution complexity of the radiotherapy plan. Among these, and The lung tissue received no less than and Volumetric parameters of irradiation.
[0071] Radiomics data are obtained based on treatment-related CT images and corresponding CTVs of the target subjects. First, the CT images and CTVs are spatially correlated, and the CTVs are identified as the volumes of interest for radiomics feature extraction.
[0072] To reduce the impact of different image spatial scales on quantitative features, CT images can be uniformly spatially processed before feature extraction. In one implementation, the CT images are isotropically resampled to achieve a voxel size of [missing information]. .
[0073] After spatial processing, the CT image is discretized with a fixed grayscale width. In one embodiment, the fixed grayscale width is set to... This process enables subsequent texture feature calculations to be performed based on a unified grayscale discretization rule.
[0074] After the above processing is completed, quantitative features are extracted from the original CT image and one or more transformed CT images, using the CTV as the volume of interest. The quantitative features include shape features, first-order statistical features, gray-level co-occurrence matrix features, gray-level dependency matrix features, gray-level run matrix features, gray-level region size matrix features, and adjacent gray-level difference matrix features.
[0075] In one experimental implementation, 1409-dimensional radiomics raw features were obtained from the CTV region. Since the dimensionality of radiomics data is significantly higher than that of clinical parameters and TP53I3 molecular pathology scoring data, this embodiment does not directly determine the 1409-dimensional radiomics raw features as the final features used in different outcome tasks. Instead, it uses them as candidate features in a unified heterogeneous multimodal feature pool, and the subsequent task-specific processing determines whether the corresponding features are retained.
[0076] TP53I3 molecular pathological scoring data is obtained based on the expression results of TP53I3 in pathological tissues. In one embodiment, the four-level expression results recorded in the pathological test results are read and encoded as 0, 1, 2, and 3 in the order of no expression, low expression, moderate expression, and high expression, respectively. After digital processing, the TP53I3 pathological test results constitute a one-dimensional molecular pathological feature.
[0077] In one implementation, the total dimension of the unified heterogeneous multimodal feature pool is 1452, including 16 dimensions of clinical parameters, 26 dimensions of radiotherapy dose parameters, 1409 dimensions of radiomics parameters, and 1 dimension of TP53I3 molecular pathology score. This results in:
[0078] For the small amount of missing data in the feature pool, the parameters used for missing value handling are determined by the corresponding training data in each model training process. Validation data and independent data outside the training phase are not involved in determining the missing value handling parameters.
[0079] For continuous variables that require scaling, the mean is also calculated using the corresponding training data. and standard deviation Standardization shall be performed according to the following formula:
[0080] in, These are the values before standardization. The values are standardized. and These are the mean and standard deviation, calculated from the training data, for the corresponding continuous variables. When processing validation or independent data, the values already generated during the training phase are directly used. and The scaling parameters are not recalculated using validation data or independent data.
[0081] In this way, data from different data sources with different dimensions and feature dimensions are first matched at the target object level, and then a unified heterogeneous multimodal feature pool is formed according to the data processing rules determined by the training data, which can be called by subsequent task branches.
[0082] Example 3 This embodiment further explains the radiotherapy response parameter generation branch based on Embodiment 1 and Embodiment 2.
[0083] For the model development data, classification labels are generated based on pre-determined radiotherapy response evaluation results. Effective responses and non-effective responses are assigned to different categories. A hierarchical five-fold cross-validation process is used to establish the training procedure, ensuring that the category composition remains approximately consistent across different cross-validation folds.
[0084] For the Each cross-validation cycle divides the current model development data into a training subset. and verification subset The two data subsets do not overlap.
[0085] For each cross-validation cycle, firstly, based solely on... Determine the parameters for handling missing values and the scaling parameters for continuous variables. After completing data processing, in Internally, the first task feature is selected. The category balancing operation also only applies to... After completing the training, use the [method / method / etc.]... Processing all data transformation rules formed .
[0086] Therefore, the validation subset does not participate in the formation of missing value handling parameters, scaling parameters, class balancing rules, or the first task feature subset. Different cross-validation cycles complete their respective data processing and model training processes based on their respective training subsets.
[0087] The first task feature screening uses features that A sparse feature selection method with regularization constraints. For candidate features in a unified heterogeneous multimodal feature pool, an optimization relationship corresponding to the classification objective is established:
[0088] in, Represents the classification loss function. Indicates the regularization strength. Indicates the first The coefficients corresponding to each candidate feature. Where, The coefficient vector that minimizes the objective function. Let be the vector of coefficients to be optimized. For category tags, For candidate feature matrices, For linear predictors, The candidate feature index is used, and the summation term is the sum of the absolute values of the coefficients of each candidate feature.
[0089] As regularization constraints are applied to candidate features, the coefficients of features that contribute less to the current radiotherapy response classification task can be compressed to 0. Features with non-zero coefficients are retained, forming the first task feature subset.
[0090] In one experimental implementation, the first task feature screening ultimately yielded 17 non-zero features. These 17 features did not originate from a single modality, but rather covered clinical parameters, radiotherapy dose parameters, radiomics parameters, and TP53I3 molecular pathology scores.
[0091] Representative features with high model gain include the percentage of predicted forced expiratory volume in one second, the percentage of predicted carbon monoxide diffusion in the lungs, CTV_RTOG Conformity Index, original_glszm_SizeZoneNonUniformityNormalized, irradiation field irregularity index, TP53I3 score, original_glcm_Autocorrelation, ECOG score, mean lung dose, and NSE.
[0092] After completing the first task feature selection, class balancing is performed within the training subset, and multiple basic classification models with different data fitting methods are trained using the resulting first task feature subset.
[0093] In one implementation, the base classification models include logistic regression, random forest, LightGBM, and XGBoost. These four base classification models can be used, or at least two can be selected to participate in subsequent ensembles.
[0094] For each base classification model, cross-validation is used to generate the out-of-bounds prediction probabilities of the training samples. For the th training samples, of which The out-of-place prediction results of the base classification model form secondary features, which serve as the training sample index. When using four basic classification models, the secondary features are represented as follows:
[0095] The corresponding training samples This serves as the training input for the two-level fusion model. Therefore, the training input for the two-level fusion model originates from the out-of-division prediction results of each base classification model for the corresponding samples, rather than the ordinary fit probabilities obtained directly by the base classification models on their own training samples. For the first The base classifier extrapolation prediction probability vector for each sample; the components of the vector correspond to the prediction probabilities of logistic regression, random forest, LightGBM, and XGBoost, respectively.
[0096] After training the basic classification model and the secondary fusion model, for the target object from which radiotherapy response parameters are to be generated, the corresponding features are first extracted based on the fixed first task feature index. Then, the already trained basic classification model is called to generate prediction results, and the multiple prediction results are combined in the same order as in the training phase to form the secondary input. Finally, the radiotherapy response probability parameters are generated by the Stacking secondary fusion model. .
[0097] Through the above processing, data preprocessing, first task feature formation, class balancing, and model training in the radiotherapy response parameter generation process are all limited to the corresponding training data. Validation data is processed only according to the rules already formed in the training data.
[0098] Example 4 This embodiment provides a lightweight radiotherapy response parameter generation method based on the integrity of the input modality.
[0099] Define the modal available states :
[0100] in, Used to indicate whether clinical parameter data meets input requirements. This is used to indicate whether the radiotherapy dose parameter data meets the input requirements. Used to indicate whether radiomics data meets the input requirements. This is used to indicate whether the TP53I3 molecular pathology scoring data meets the input requirements. This represents a vector consisting of the available input states for the four modalities; each component is represented by 1 to indicate that the corresponding modality can be used as model input, and by 0 to indicate that the corresponding modality does not meet the preset input requirements.
[0101] When all four modalities meet the input requirements, a complete unified heterogeneous multimodal feature pool is constructed in accordance with the method described in Example 1, and complete task adaptive processing is performed according to the task type.
[0102] In actual data acquisition, some target subjects may not be able to obtain complete radiotherapy dose parameters or radiomics data simultaneously. For radiotherapy response parameter generation tasks, when at least one of the radiotherapy dose parameter data or radiomics data is unavailable, but the clinical parameter data and TP53I3 molecular pathology score data meet the input requirements, a lightweight radiotherapy response processing pathway can be adopted.
[0103] In this lightweight processing pathway, clinical parameter data is... Comparison with TP53I3 molecular pathology scoring data Perform case-level matching to form a lightweight feature set:
[0104] The data processing rules for lightweight feature sets, as well as the feature indices and model parameters used, are determined during the training phase of the corresponding lightweight model. For new target objects, the data processing parameters fixed during the training phase are directly adopted. The data transformation is performed, and then a pre-trained lightweight radiotherapy response model is invoked to generate radiotherapy response probability parameters. This represents a lightweight feature vector formed by clinical parameters and TP53I3 scores of the same case in a predetermined order.
[0105] This lightweight processing method does not require temporarily deleting input features from the full model after data loss. Instead, it calls a lightweight model pre-trained based on clinical parameters and the TP53I3 molecular pathology score. Therefore, its feature processing parameters and model parameters correspond to this input modality combination.
[0106] In one experimental dataset, radiotherapy response parameters were generated using a low-dimensional combination of clinical parameters and TP53I3 molecular pathology scores. The five-fold cross-validation AUC was approximately [missing value]. .
[0107] Compared to the complete processing path that requires CTV radiomics feature calculation and radiotherapy planning data parsing, this implementation reduces the reliance on the aforementioned data processing procedures when the input modality is incomplete. The lightweight path in this embodiment is used for generating radiotherapy response probability parameters.
[0108] Example 5 This embodiment further explains the process of generating progression-free survival risk parameters based on Embodiment 1.
[0109] For case data entering the progression-free survival task branch, the progression-free survival time of the target subjects is also recorded. and event status When a preset progress event occurs during the observation period, set... If the aforementioned progress event has not occurred by the end of the observation period, set... .
[0110] Unify heterogeneous multimodal features No-progression survival time and event status Together they form censored live data, in which For the first Multimodal features of a target object The no-progression survival time of the target object. For event status, Index representing the target object:
[0111] First, a Cox survival model with regularization constraints is used to process candidate features in a unified heterogeneous multimodal feature pool. The risk relationship is represented as:
[0112] in, Representing a given feature In time The risk function, Represents the baseline risk function. Indicates the observation time. This represents the survival association parameter vector corresponding to the candidate feature. This represents the transpose of the vector. This represents an exponential function.
[0113] Candidate features are ranked based on regularization path, non-zero coefficients, or feature importance to form a ranked candidate feature set. ,in This represents the total number of candidate features. The top-ranked features are selected from the ranking results. The candidate features form a subset of the second task features. .
[0114] In one implementation, ,in The number of candidate features selected. The top 50 features in the resulting ranking are mainly CTV radiomics features, and also include radiotherapy dose parameters and clinical parameters.
[0115] The first and second task feature subsets were formed based on different supervision objectives. In the experimental data, the radiotherapy response task generated 17 task-specific features, while the progression-free survival task generated the top 50 survival-related features. The number of features, feature composition, and contributions from different modalities can all differ between the two feature subsets.
[0116] Subset of features of the second task Input the random survival forest model and the gradient boosting survival analysis model respectively. The random survival forest model outputs risk parameters. The gradient boosting survival analysis model outputs risk parameters. .
[0117] The two risk parameters are then fused according to predetermined weights:
[0118] in, The weights represent the fusion weights of the random survival forest model and the weights of the gradient boosting survival analysis model. , satisfy:
[0119] In one implementation:
[0120] Therefore, we can conclude that:
[0121] The fusion weights are predetermined based on model performance in the training data and remain fixed during the independent deployment phase after training is completed.
[0122] In obtaining progression-free survival risk parameters Subsequently, based on the risk thresholds or risk parameter quantiles determined during the training phase, continuous risk parameters can be mapped to different risk levels. For example, the risk parameters corresponding to the target object can be mapped to different risk levels for subsequent data display, sorting, or statistical analysis by the information system.
[0123] The risk level is obtained by post-processing the model output parameters according to preset rules. The data processing does not change the training parameters of the aforementioned RSF model, GBSA model, and risk fusion model.
[0124] Example 6 This embodiment further illustrates the process of deploying and calling the training model on an independent target object.
[0125] After the model training phase is completed, the data processing information required for generating radiotherapy outcome parameters is saved as model configuration data. This model configuration data includes parameters for handling missing values, scaling, task feature indexes, classification model parameters, Stacking fusion parameters, survival analysis model parameters, and risk fusion parameters.
[0126] For the radiotherapy response task, the model configuration data should at least include the feature index of the corresponding first task feature subset, as well as the parameters of the basic classification model and the second-level fusion model that participate in the calculation of the radiotherapy response probability.
[0127] For no-progression survival tasks, the model configuration data should include at least the feature index of the corresponding second task feature subset, as well as the RSF model, GBSA model, and corresponding risk fusion weights.
[0128] For new target objects outside the training data, case-level mappings are first established between different modalities based on the target object identifier. Subsequently, missing value handling, feature encoding, and data scaling are performed according to the data processing rules corresponding to the invoked model.
[0129] When dealing with continuous variables, the mean that has been fixed during the training phase can be directly called. and standard deviation Complete the Z-score transformation. The data from the new target object will not be used for re-determination. or .
[0130] After the data transformation is complete, the input features actually needed by the model are selected from the unified heterogeneous multimodal data based on the task feature index corresponding to the current parameters to be generated. This process does not re-perform task feature filtering for new target objects.
[0131] When the parameter to be generated is the radiotherapy response probability parameter, the selected first task feature is sequentially input into multiple already trained basic classification models, and then the radiotherapy response probability parameter is obtained through the Stacking secondary fusion model.
[0132] When the parameter to be generated is the no-progression survival risk parameter, the selected second task feature is input into the already trained RSF model and GBSA model respectively, and then the parameters fixed during the training phase are used. Complete risk integration and obtain .
[0133] Therefore, the new target object only performs the already determined data transformation, feature selection index invocation, and model forward computation, and does not participate in missing value handling parameter re-estimation, scale parameter re-estimation, task feature re-selection, model refitting, or fusion weight redeter determination.
[0134] In one implementation, an independent external cohort of 90 patients who did not participate in model training, feature selection, and parameter optimization was formed. Clinical parameters, DVH parameters, CTV radiomics data, and TP53I3 molecular pathology scores of this independent external cohort were matched at the case level, and the data processing parameters, task feature indexes, and model configuration data already fixed during the training phase were fully reused.
[0135] The AUC of the radiotherapy response model in the independent external cohort is approximately The accuracy rate is approximately , The fraction is approximately Sensitivity is approximately The specificity is approximately .
[0136] The C-index of the no-progression survival risk fusion model in the independent external cohort is approximately After mapping risk levels according to predetermined risk parameters, the Kaplan-Meier survival curves corresponding to different risk levels become separated.
[0137] Therefore, the data transformation parameters, task feature indexes, model parameters, and fusion parameters formed during the model development phase can remain consistent in independent data processing, avoiding the re-entry of independent data into the training parameter formation process, and enabling a unified data processing link between the training phase and the independent deployment phase.
[0138] The technical solutions of this application have been described above with reference to specific embodiments. It should be understood that the above embodiments are used to illustrate the technical solutions of this application and are not intended to limit the scope of protection of this application. Without departing from the technical concept of this application, those skilled in the art can make corresponding equivalent substitutions or adaptive adjustments to the data acquisition, data processing, model calling, and parameter output processes based on the actual data source, data completeness, and computing environment. Any technical solution based on the technical content disclosed in this application, that forms a corresponding task feature subset according to the data structure of different radiotherapy outcome tasks on the basis of unified heterogeneous multimodal data, and performs corresponding parameter generation processing, should have its protection boundary determined according to the scope defined by the claims of this application.
[0139] The technical features described in the various embodiments of this application can be combined with each other where there is no technical contradiction. The descriptions of specific data dimensions, number of features, number of training samples, model weights, and verification results in each embodiment are used to illustrate the specific implementation of the corresponding technical solutions. Except for the content explicitly defined in the claims, the above specific values should not be construed as necessary limitations on other embodiments.
[0140] The computer implementation process of this application can be executed by a computer device with data processing capabilities. The data processing parameters, task feature indexes, model parameters, and fusion parameters generated after model training can be stored as model configuration data and invoked by the processor when processing new target objects to perform corresponding data transformation, feature extraction, and forward computation processes. The above-described computer implementation is a specific execution of the method steps of this application and does not change the technical concept of generating lung cancer radiotherapy outcome parameters based on task-adaptive heterogeneous multimodal fusion.
Claims
1. A method for generating radiotherapy outcome parameters for lung cancer based on task-adaptive heterogeneous multimodal fusion, characterized in that, include: Acquire clinical parameters, radiotherapy dose parameters, CT radiomics parameters with the clinical target volume as the volume of interest, and TP53I3 molecular pathology score for the same patient; Match four-modal data according to patient identifiers, complete missing value and data scaling processing, and construct a unified heterogeneous multimodal feature pool under the same case index; Adaptive routing is performed based on the outcome parameter type, calling either the radiotherapy response classification branch or the progression-free survival analysis branch respectively. Radiation therapy response branch, based on response class labels, utilizes... The sparse screening of regularization constraints retains features with non-zero coefficients, forming a subset of features for the first task. Class balancing is performed within the training subset, and the prediction results of at least two classification models are integrated to generate radiotherapy response probability parameters. The no-progression survival branch sorts and filters candidate features based on survival time and progress event status using a Cox survival model with regularization constraints, forming a subset of features for the second task. By fusing the risk outputs of at least two survival analysis models, a progression-free survival risk parameter is generated. After training, missing values, scale parameters, task feature indexes, and model parameters are fixed; these parameters are used for new patients without re-selecting features, and external data is not used for model refitting. Output at least one of a radiotherapy response probability parameter and a progression-free survival risk parameter; wherein the two task feature subsets are formed based on a classification objective and a censored survival objective, respectively, and differ in at least one of modality composition, specific features, or feature weights.
2. The method for generating lung cancer radiotherapy outcome parameters by task-adaptive heterogeneous multimodal fusion according to claim 1, characterized in that, Missing value processing and feature encoding are performed on each modality of data, with missing value processing parameters determined by the training data; the mean of continuous variables is determined based on the training data. and standard deviation ,according to Scale processing is performed, among which, To process the eigenvalues beforehand, For the processed eigenvalues, The mean of the training data for the aforementioned feature. The standard deviation of the training data for the aforementioned feature is used; the corresponding processing parameters are fixed and applied to the data to be processed outside of the training data.
3. The method for generating lung cancer radiotherapy outcome parameters by task-adaptive heterogeneous multimodal fusion according to claim 1, characterized in that, The clinical parameter data include at least one of the following parameters: age, sex, smoking status, pathological type, ECOG score, percentage of predicted forced expiratory volume in one second, percentage of predicted carbon monoxide diffusion capacity in the lungs, TNM stage, NSE, CYFRA21-1, CEA, SCC, and CA153.
4. The method for generating lung cancer radiotherapy outcome parameters by task-adaptive heterogeneous multimodal fusion according to claim 1, characterized in that, The radiotherapy dose parameter data includes at least one of the following parameters: fractionation parameter, beam parameter, treatment plan complexity parameter, minimum dose to clinical target area, maximum dose to clinical target area, average dose to clinical target area, conformity index, homogeneity index, dose gradient index, average lung dose, and lung tissue dose. and lung tissue The and Each characterizes lung tissue receiving at least and Volumetric parameters of irradiation.
5. The method for generating lung cancer radiotherapy outcome parameters by task-adaptive heterogeneous multimodal fusion according to claim 1, characterized in that, The acquisition of the radiomics data includes: spatial resampling and fixed gray-width discretization of CT images using the clinical target area as the volume of interest; and extracting shape features, first-order statistical features, gray-level co-occurrence matrix features, gray-level dependency matrix features, gray-level run-length matrix features, gray-level region size matrix features, and adjacent gray-level difference matrix features from the original CT images and the transformed CT images; wherein, the voxel size of the spatial resampling is... Fixed grayscale width is .
6. The method for generating lung cancer radiotherapy outcome parameters by task-adaptive heterogeneous multimodal fusion according to claim 1, characterized in that, The TP53I3 molecular pathological scoring data is obtained by digitally encoding the four-level expression results of TP53I3 recorded in the pathological test results. The four-level expression results are no expression, low expression, moderate expression and high expression, and are encoded as 0, 1, 2 and 3, respectively.
7. The method for generating lung cancer radiotherapy outcome parameters by task-adaptive heterogeneous multimodal fusion according to claim 2, characterized in that, In the radiotherapy response task, stratified five-fold cross-validation was employed. In each cross-validation cycle, missing value handling parameters and scaling parameters were determined within the training subset. [The last sentence appears to be incomplete and possibly refers to a different context.] The sparse feature selection method with regularization constraints retains the features corresponding to non-zero coefficients to form the first task feature subset, and performs class balancing and classification model training by synthesizing minority class oversampling method; the corresponding validation subset does not participate in parameter determination, feature selection and class balancing.
8. The method for generating lung cancer radiotherapy outcome parameters by task-adaptive heterogeneous multimodal fusion according to claim 7, characterized in that, The at least two classification models are selected from logistic regression, random forest, LightGBM, and XGBoost models. Cross-validation is used to obtain the out-of-division prediction probabilities of each basic classification model for training samples that did not participate in the current fitting. The out-of-division prediction probabilities of the same training sample are combined into a secondary input according to the model order. The secondary input is processed by the Stacking secondary fusion model to generate radiotherapy response probability parameters.
9. The method for generating lung cancer radiotherapy outcome parameters by task-adaptive heterogeneous multimodal fusion according to claim 1, characterized in that, In the no-progression survival task, a Cox survival model with regularization constraints is adopted. Features are ranked based on their association with no-progression survival time and progress event states, and the top-ranked features are selected. The candidate features form a subset of the second task features; among them... This indicates the number of features selected, and .
10. The method for generating lung cancer radiotherapy outcome parameters by task-adaptive heterogeneous multimodal fusion according to claim 9, characterized in that, The second task feature subset is input into the random survival forest model and the gradient boosting survival analysis model, respectively, to obtain the risk output. and ,according to Weighted fusion is performed, in which, For progression-free survival risk parameters, For the fusion weights of the random survival forest model, and .
Citation Information
Patent Citations
Recurrent nasopharyngeal carcinoma radiotherapy nasopharyngeal necrosis prediction method
CN115564735A
Non-small cell lung cancer prognosis prediction method based on multi-modal fusion
CN118314106A
