CN-122025190-A - Multi-source collaborative dynamic time sequence TCP prediction method, system, medium and program product

CN122025190ACN 122025190 ACN122025190 ACN 122025190ACN-122025190-A

Abstract

The invention discloses a multisource collaborative dynamic time sequence TCP prediction method, a system, a medium and a program product, wherein the method comprises the steps of collecting multisource heterogeneous data of a patient radiotherapy full period, wherein the multisource heterogeneous data comprises DVH data, CT data, CPP data and longitudinal follow-up data of a plurality of time nodes; the method comprises the steps of carrying out layered preprocessing and data format unification on multi-source heterogeneous data, constructing a dynamic time sequence feature fusion matrix containing time dependency, constructing a deep learning model integrating a time sequence processing unit and a cross-mode fusion unit, supporting an incremental learning iterative updating mechanism by the model, outputting TCP prediction results of a new patient at different time nodes by using the model, calling a clinical rule adaptation model if the new patient is a special case, and correcting the TCP prediction results by adopting a mixed correction strategy combining rule matching and doctor experience weight. The method can deeply fuse multi-source heterogeneous data, capture time dynamic association, quantify the contribution degree of modal characteristics, support continuous iterative update and adapt to individual differences.

Inventors

ZENG LEI
OUYANG HUIDAN
CHEN SHUYUN
Xiang Lele
DENG JIALE

Assignees

南昌大学第二附属医院

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. A multi-source collaborative dynamic time sequence tumor control probability TCP prediction method is characterized by comprising the following steps: acquiring DVH data, CT data and CPP data before radiotherapy of a patient, and longitudinal follow-up data of tumor control states and physiological index change data of a plurality of time nodes after radiotherapy; Carrying out layered preprocessing and data format unification on the collected multi-source heterogeneous data; Quantifying the contribution rate of each modal feature to the TCP prediction result through ScaledDot-Product attention mechanism, realizing cross-modal feature alignment and preliminary fusion, and constructing a dynamic time sequence feature fusion matrix containing time dependency relationship by taking a longitudinal follow-up time axis as an index, associating the preliminary fusion feature with tumor control state and physiological index change data of a corresponding time node; constructing a deep learning model integrating a time sequence processing unit and a cross-modal fusion unit based on the matrix, wherein the model adopts a dynamic weight mixing loss function, and the model parameters are optimized through AdamW to support an incremental learning iterative updating mechanism; And if the new patient is a special case, calling a clinical rule adaptation model, and adopting a mixed correction strategy combining rule matching and doctor experience weight to correct the TCP prediction result.
2. The method of claim 1, wherein, The DVH data specifically comprises a dose distribution parameter and an irradiated volume parameter; The CPP data specifically comprises patient age, tumor TNM stage, tumor marker level, primary tumor size, pathological type, differentiation degree, whether key basic diseases are combined or not, and whether combined treatment is received or not; The CT data specifically comprises tumor density distribution data of a CT image and tumor texture feature data in T1 and T2 weighted images of MRI, wherein the tumor density distribution data of the CT image specifically comprises an average CT value and a density standard deviation, and the tumor texture feature data in the T1 and T2 weighted images of MRI specifically comprises gray level co-occurrence matrix entropy, contrast and energy; The multiple time nodes of the longitudinal follow-up data are set through a self-adaptive time sequence follow-up mechanism, specifically, 1 time per month follow-up nodes are set for patients 1-3 months after radiotherapy and used for capturing early tumor response conditions, 1 time per 3 months follow-up nodes are set for patients 3-24 months after radiotherapy and used for tracking the medium-term tumor control effect, 1 time per 6 months follow-up nodes are set for patients 24-60 months after radiotherapy and used for tracking the long-term tumor control effect, and 1 time per year follow-up nodes are set for patients more than 60 months after radiotherapy and used for model iterative updating without including core prediction time nodes; the tumor control state is clearly divided into four types of complete alleviation, partial alleviation, stability and progress according to RECIST; the physiological index change data specifically comprises blood routine index, hypersensitive C-reactive protein and plasma EBV-DNA copy number; The dynamic weight mixing loss function specifically comprises cross entropy loss and time sequence mean square error loss, and the formula is as follows: , In the above-mentioned method, the step of, Representing the dynamic weight mixing loss of the dynamic weight, In order for the cross-entropy loss to occur, For the time-series mean square error loss, Is a dynamic weight coefficient; The special clinical conditions corresponding to the special cases comprise more than or equal to 1/10000 of rarity of tumor pathological types, combination of more than 2 severe basic diseases, more than or equal to 15% of target area dose deviation of radiotherapy plan standard scheme, and combination treatment including oncolytic virus treatment, antibody coupled drug targeting treatment, cell treatment or gene treatment.
3. The method according to claim 1, wherein the step of hierarchical preprocessing specifically comprises: adopting a dose quantile normalization method for the DVH data to eliminate dose scale deviation; Automatically extracting the edge contour definition, internal density distribution and internal heterogeneity characteristics of the tumor from the CT data through MobileNet, and simultaneously completing gray scale normalization and size unification; and carrying out abnormal value detection elimination, missing value completion and data standardization processing on the CPP data and the longitudinal follow-up data in sequence.
4. The method according to claim 1, wherein the step of unifying the data formats is specifically: And finally uniformly converting all the multi-modal characteristics after hierarchical preprocessing into a tensor format of an adaptive deep learning framework, wherein the space-time variation characteristics of the CT data are stored in 4-dimensional tensors, the number of dimensions is the number of samples multiplied by the height multiplied by the width multiplied by the channel number, the dose-volume correlation characteristics of the DVH data, the static characteristics of the CPP data and the follow-up characteristics of the longitudinal follow-up data are stored in 2-dimensional tensors, and the number of dimensions is the number of samples multiplied by the number of characteristics.
5. The method of claim 1, wherein the core calculation formula of ScaledDot-Product attention mechanism is: , In the above-mentioned method, the step of, To correspond to the query vector of target features to be matched, For a key vector corresponding to a reference feature to be retrieved, To correspond to a vector of values for which the output characteristics are to be weighted, Is based on 、、 The calculated attention weighted fusion features, For the normalization of the exponential function, As the dimension of the key vector, Is a transpose of the matrix.
6. The method of claim 1, wherein the architecture of the deep learning model is as follows: The time sequence processing unit adopts a cascade structure of Bi-LSTM and a transducer encoder, wherein the Bi-LSTM is provided with 2 hidden layers, the dimension of the hidden layers is 256, the transducer encoder comprises 4 encoding layers, and each encoding layer is provided with 8 attention heads; the cross-modal fusion unit dynamically distributes weights of single-modal feature vectors corresponding to DVH data, CT data and CPP data by adopting a gating mechanism, wherein a gating weight formula specifically comprises the following steps: , In the above-mentioned method, the step of, In order to gate the weights on the basis of the weight, The function is activated for Sigmoid, As a matrix of weights, the weight matrix, Respectively, the single-mode feature vectors of DVH data, CT data and CPP data after cross-mode alignment, Is a bias term; And the output layer of the deep learning model outputs the TCP corresponding to each time node through the full connection layer and the Softmax activation function.
7. The method of claim 1, wherein the cross-modal feature is dynamically feature adapted by MMD algorithm to match the feature distribution of the new patient to the feature distribution of the model training data, the dynamic feature adaptation being formulated as follows: , In the above-mentioned method, the step of, To train data characteristics, For the characteristics of the data of the new patient, To be used for measuring And The maximum mean difference in distance between the two probability distributions, As a function of the gaussian kernel, Is the first The characteristics of the individual training data are such that, Is the first The characteristics of the individual training data are such that, Is the first The data characteristics of the new patient are provided, Is the first The data characteristics of the new patient are provided, For the total number of training data features, Is the total number of data features for the new patient.
8. A multi-source coordinated dynamic time-series tumor control probability TCP prediction system, comprising: The multi-component heterogeneous data acquisition module is used for acquiring DVH data, CT data and CPP data before radiotherapy of a patient and longitudinal follow-up data of tumor control states and physiological index change data of a plurality of time nodes after radiotherapy; The pre-processing module is used for carrying out layered preprocessing and data format unification on the collected multi-source heterogeneous data; The dynamic time sequence feature fusion module is used for quantifying the contribution rate of each modal feature to the TCP prediction result through a ScaledDot-Product attention mechanism, realizing cross-modal feature alignment and preliminary fusion, and constructing a dynamic time sequence feature fusion matrix containing time dependence by taking a longitudinal follow-up time axis as an index, associating the preliminary fusion feature with tumor control state and physiological index change data of a corresponding time node; The multi-unit collaborative model training module is used for constructing a deep learning model integrating a time sequence processing unit and a cross-modal fusion unit based on the matrix, wherein the model adopts a dynamic weight mixing loss function, and model parameters are optimized through AdamW to support an incremental learning iterative updating mechanism; And the personalized dynamic prediction module is used for inputting the multisource heterogeneous data of the new patient before radiotherapy into the model after preprocessing and dynamic characteristic adaptation, outputting TCP prediction results of different time nodes, and calling a clinical rule adaptation model and adopting a mixed correction strategy combining rule matching and doctor experience weight to correct the TCP prediction results if the new patient is a special case.
9. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any of claims 1-7.
10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any of claims 1-7.

Description

Multi-source collaborative dynamic time sequence TCP prediction method, system, medium and program product Technical Field The invention relates to the technical field of medical data intelligent analysis and tumor radiotherapy, in particular to a multisource collaborative dynamic time sequence TCP (Tumor Control Probability ) prediction method, a system, a medium and a program product. Background Radiation therapy is one of the core means of local treatment of tumors, and is widely applied to clinical diagnosis and treatment of various tumors such as nasopharyngeal carcinoma, breast cancer and the like, and the key index of curative effect evaluation is TCP (transmission control protocol), namely the probability of tumor regression or stabilization within a specific time period after radiation therapy. In clinical practice, accurate evaluation of TCP has important significance for individual radiotherapy scheme formulation, follow-up strategy optimization and patient prognosis judgment, and is also a core requirement to be improved in the current tumor radiotherapy field. Aiming at the requirement of TCP evaluation, the prior art mainly adopts two types, namely, one type is to rely on a traditional empirical formula, such as Niemierko model, and calculate by quantifying the association of radiotherapy dose distribution and tumor control, and the other type is to adopt a simple statistical model, such as logistic regression, random forest and the like, and complete TCP prediction by combining a small amount of clinical or dose data. However, the above prior art has a plurality of defects in practical application, and is difficult to meet the requirements of clinical precision and practicality, and the specific steps are as follows: 1. the data utilization is single, the synergistic value of the multi-source heterogeneous information is not mined, the prior art only depends on DVH (Dose-Volume Histogram) data or single clinical indexes (such as tumor stage), and the synergistic association of tumor image characteristics (such as CT density distribution and MRI texture heterogeneity) and longitudinal follow-up physiological indexes (such as plasma EBV-DNA copy number) and Dose data is ignored, so that the predicted dimension is one-sided. 2. The time dynamic effect is lost, the long-term control and evaluation is insufficient, the existing model is mostly static prediction, only TCP at a single time point (such as 6 months after radiotherapy) is output, the dynamic change of TCP at a plurality of time points after radiotherapy cannot be captured, and the requirements of clinical long-term follow-up and scheme adjustment are difficult to meet. 3. The cross-modal data fusion is low-efficiency, the prediction precision is limited, the data type heterogeneity exists in DVH data (dose-volume continuous value), image data (high-dimensional pixel characteristics) and clinical data (discrete classification/continuous value), the existing model lacks an effective fusion mechanism, the contribution degree of different modal characteristics to a prediction result cannot be quantized, and key characteristics are easy to lose. 4. The model updating mechanism is missing, the clinical adaptability is attenuated, the historical training data distribution can be dynamically changed along with the accumulation of new cases and the iteration of radiotherapy technology (such as proton radiotherapy and stereotactic radiotherapy), but the traditional model cannot be incrementally incorporated into new data, the full model is required to be retrained, the operation is tedious, and the time-dependent requirement of clinical data is difficult to adapt. 5. The individuation suitability is insufficient, the prediction deviation of special cases is large, and the existing model does not design a special correction strategy aiming at special clinical situations such as rare pathological types, combined severe basic diseases and the like, so that the prediction result and the actual deviation of the cases are large. Aiming at the technical problems, a TCP prediction scheme which can deeply fuse multi-source heterogeneous data, dynamically correlate capturing time, quantify modal characteristic contribution degree, support continuous iteration update and adapt individual difference is needed to improve accuracy and clinical practicality of radiotherapy curative effect evaluation, and scientific support is provided for personalized radiotherapy scheme formulation and follow-up strategy optimization. Disclosure of Invention The invention provides a multisource collaborative dynamic time sequence TCP prediction method, a system, a medium and a program product, which solve the problems of single data utilization, time dynamic effect deletion, cross-modal data fusion inefficiency, model updating mechanism deletion and insufficient individuation suitability in the prior art. In a first aspect, the present invention provides a method for predicti