Lung adenocarcinoma postoperative progress prediction method

Abstract

The invention discloses a lung adenocarcinoma postoperative progress prediction method, which comprises the following steps of: firstly, acquiring a lung adenocarcinoma postoperative clinical data set and CT (Computed Tomography) image data, and the clinical data comprises a plurality of clinical parameters and postoperative condition labels; a clinical data model is constructed based on the clinical data set, and an independent predictive factor is obtained based on the model. A radiomics feature model is constructed according to CT image data, and a feature weight is calculated based on the model. And constructing an image feature extraction model according to the CT image data and the pre-training model, and obtaining a learning feature score based on the model. And integrating the independent prediction factor, the feature weight and the learning feature score to construct a target prediction model. In practical application, the prediction result can be obtained by inputting the lung adenocarcinoma postoperative data to be predicted. According to the embodiment of the invention, by integrating multi-source data modeling, the limitation of a single data source can be broken through, so that the accuracy of postoperative prediction of lung adenocarcinoma is effectively improved.

Description

Technical Field

[0001] The embodiments of the present application relate to, but are not limited to, the field of surgical technology, and in particular to a method for predicting the postoperative progression of lung adenocarcinoma. Background Art

[0002] Current methods for predicting the postoperative progression of lung adenocarcinoma have many limitations: traditional clinical data prediction relies on markers, but single markers are subject to interference from multiple factors, have poor stability and reliability, and different research conclusions vary greatly, making it difficult to form a unified standard. For predictions based on CT images, manual analysis of imaging features is subject to large differences in subjective judgment, and only partial morphological features of the tumor can be obtained, which cannot fully reflect the biological behavior of the tumor. Although deep learning has been applied in this field and can extract features from images, it is highly dependent on large amounts of high-quality data, making it difficult to ensure the accuracy and consistency of data annotation, and the model has poor interpretability. Existing prediction methods are mostly based on a single data source, making it difficult to integrate multi-dimensional information. There is also a lack of effective fusion between different methods, making it difficult to obtain accurate prediction results. Summary of the Invention

[0003] The following is a summary of the subject matter described in detail herein. This summary is not intended to limit the scope of the claims.

[0004] The present invention provides a method for predicting the postoperative progression of lung adenocarcinoma, which can effectively improve the accuracy of postoperative prediction of lung adenocarcinoma.

[0005] An embodiment of the present application provides a method for predicting the postoperative progression of lung adenocarcinoma, comprising: obtaining a clinical data set and CT imaging data after surgery for lung adenocarcinoma, the clinical data set comprising multiple clinical parameters and postoperative condition labels; constructing a clinical data model based on the clinical data set, and obtaining an independent predictor based on the clinical data model, wherein the independent predictor is a parameter among the multiple clinical parameters that has a predictive effect on the dependent variable of the clinical data model; constructing an imaging genomics feature model based on the CT imaging data, and obtaining a feature weight based on the imaging genomics feature model; constructing an image feature extraction model based on the CT imaging data and a pre-trained model, and obtaining a learning feature score based on the image feature extraction model; constructing a target prediction model based on the independent predictor, the feature weight, and the learning feature score; and inputting the postoperative data of lung adenocarcinoma to be predicted into the target prediction model to obtain a prediction result.

[0006] In one embodiment of the present application, a clinical data model is constructed based on the clinical data set, and independent predictors are obtained based on the clinical data model, including: modeling according to each of the clinical parameters and the postoperative condition labels in the clinical data set to obtain a logistic regression model of each of the clinical parameters; determining a relationship value between the postoperative condition label and each of the clinical parameters based on the logistic regression model of each of the clinical parameters; screening out the clinical parameters whose relationship values ​​are less than a preset threshold from the clinical data set to obtain a predictor set; and confirming independent predictors from the predictor set.

[0007] In one embodiment of the present application, confirming an independent predictor from the predictor set includes: constructing a first clinical data model based on the postoperative condition label; adding the predictor set to the first clinical data model, and calculating a first meaning relationship value between the predictor set and the postoperative condition label; screening out the predictor set corresponding to a value less than a preset threshold value, to obtain a first initial independent predictor; modeling based on the predictor set and the postoperative condition label to obtain a second clinical data model; calculating a second meaning relationship value corresponding to each of the predictor sets in the second clinical data model, and screening out the predictor set corresponding to a value less than the preset threshold value, until in the second clinical data model, the second meaning relationship value corresponding to each of the predictor sets is less than the preset threshold value, to obtain a second initial independent predictor; confirming an independent predictor based on the first initial independent predictor and the second initial independent predictor.

[0008] In one embodiment of the present application, constructing an imaging genomics feature model based on the CT image data includes: marking the lesion area of ​​each CT image in the CT image data to obtain a key area image corresponding to each CT image; performing feature extraction on the key area image to obtain an imaging feature data set; performing feature screening processing on the imaging feature data set to obtain a modeling feature data set; and constructing an imaging genomics feature model based on the modeling feature data set.

[0009] In one embodiment of the present application, obtaining feature weights based on the imaging genomics feature model includes: extracting the regression coefficient of the modeling feature data set from the imaging genomics feature model; and calculating based on the absolute value of the regression coefficient to obtain the feature weights of each feature.

[0010] In one embodiment of the present application, the feature screening processing of the image feature dataset to obtain a modeling feature dataset includes: screening the image feature dataset to obtain a preliminary feature dataset; performing a one-way analysis of variance on the preliminary feature dataset to obtain an intermediate processed feature dataset; performing a regression analysis on the intermediate processed feature dataset, and screening the modeling feature dataset based on the analysis results.

[0011] In one embodiment of the present application, the image feature extraction model is constructed based on the CT image data and the pre-trained model, including: resampling, lesion area annotation and stretching processing of all CT images in the CT image data to obtain target processed images corresponding to each of the CT images; adjusting parameters of the pre-trained model based on the CT image data to obtain a target feature extraction model; using the target feature extraction model, performing feature extraction and feature screening processing on the target processed images corresponding to each of the CT images to obtain target image features; and constructing an image feature extraction model based on the target image features.

[0012] In one embodiment of the present application, the target feature extraction model is used to perform feature extraction and feature screening on the target processed images corresponding to each of the CT images to obtain target image features, including: merging the target processed images corresponding to the CT images to obtain multiple target image blocks, each of the target image blocks including a first processed image, a second processed image, and a third processed image, wherein the first processed image corresponds to a first color channel, the second processed image corresponds to a second color channel, and the third processed image corresponds to a third color channel; using the target feature extraction model, performing feature extraction on the target image blocks to obtain preliminary image feature data; dividing the preliminary image feature data into channels according to the first color channel, the second color channel, and the third color channel to obtain first channel feature data, second channel feature data, and third channel feature data; extracting a first target feature from the first channel feature data, extracting a second target feature from the second channel feature data, and extracting a third target feature from the third channel feature data to obtain target channel feature data; performing feature screening on each feature in the target channel feature data to obtain intermediate image feature data; and performing feature screening on the intermediate image feature data using ten-fold cross validation to obtain target image features.

[0013] In one embodiment of the present application, the learning feature score is obtained based on the image feature extraction model, including: obtaining a regression coefficient set of the image feature extraction model; determining the regression coefficient of each feature in the target image feature based on the regression coefficient set; and calculating the learning feature score based on the regression coefficient of each feature.

[0014] In one embodiment of the present application, the target prediction model is constructed based on the independent prediction factors, the feature weights and the learned feature scores, including: standardizing the independent prediction factors and the learned feature scores, and constructing an initial prediction model based on the processed independent prediction factors, the learned feature scores and the feature weights; obtaining a prediction training set and a prediction test set; training the initial prediction model based on the prediction training set, and adjusting the weight parameters of the initial prediction model based on a preset error index, and obtaining the initial prediction model when the maximum number of iterations is reached or the error converges to a preset threshold; evaluating the initial prediction model using the prediction test set, and when the performance of the initial prediction model in the prediction test set meets the requirements, using the initial prediction model as the target prediction model.

[0015] The embodiment of the present application provides a method for predicting the postoperative progression of lung adenocarcinoma. First, a clinical data set and CT imaging data of a patient with lung adenocarcinoma after surgery are obtained. The clinical data include multiple clinical parameters and corresponding postoperative condition labels. Then, a clinical data model is constructed based on the collected clinical data set. By analyzing the model, parameters with significant predictive effects on the clinical data model are screened out from multiple clinical parameters. These parameters are independent predictors. Next, an imaging genomics feature model is constructed for the CT imaging data, and feature weights are obtained based on the imaging genomics feature model. Subsequently, an image feature extraction model is constructed using the CT imaging data and the pre-trained model. Through this model, a learning feature score reflecting the postoperative characteristics of lung adenocarcinoma can be further obtained. After obtaining the independent predictors, feature weights, and learning feature scores, a target prediction model is constructed based on these data. In actual application, the postoperative data of lung adenocarcinoma to be predicted is input into the target prediction model to obtain the prediction results of the postoperative progression of lung adenocarcinoma. The embodiment of the present application integrates independent predictive factors obtained based on a clinical data model, feature weights obtained based on an imaging genomics feature model, and learning feature scores obtained based on an image feature extraction model to build a model, surpassing the performance of a single data source model, thereby effectively improving the accuracy of postoperative prediction of lung adenocarcinoma. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a method for predicting lung adenocarcinoma postoperative progression provided in an embodiment of the present application;

[0017] Figure 2 This embodiment of the present application provides Figure 1 Specific flow chart of step 120;

[0018] Figure 3 This embodiment of the present application provides Figure 2 Specific flow chart of step 240;

[0019] Figure 4 This embodiment of the present application provides Figure 1 Specific flow chart of step 130;

[0020] Figure 5 is an original image of a CT image provided in a specific example of this application;

[0021] Figure 6 It is a key area image provided by a specific example of this application;

[0022] Figure 7 This embodiment of the present application provides Figure 1 Specific flow chart of step 140;

[0023] Figure 8 This embodiment of the present application provides Figure 7 Specific flow chart of step 730;

[0024] Figure 9 This embodiment of the present application provides Figure 1 Specific flow chart of step 150 in FIG. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0026] It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in an order different from that in the flowchart. The terms "first" and "second" in the specification, claims, and the above-mentioned figures are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the content disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limitations of the implementation of this application. Therefore, they have no technical significance. Any modification of the structure, change in the proportional relationship, or adjustment of the size should still fall within the scope of the technical content disclosed in this application without affecting the efficacy and purpose of this application. At the same time, the terms such as "upper", "lower", "left", "right", "middle", and "one" cited in this specification are only for the convenience of description and are not used to limit the scope of the implementation of this application. Changes or adjustments in their relative relationships should also be considered as the scope of the implementation of this application without substantially changing the technical content.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0028] Current methods for predicting postoperative progression of lung adenocarcinoma suffer from numerous limitations. Traditional clinical data prediction relies on biomarkers such as insulin-like growth factor receptor-1 (IGFR1). However, single markers are subject to interference from multiple factors, resulting in poor stability and reliability. Conclusions vary widely between studies, making it difficult to establish a unified standard. Prediction based on CT images, when analyzing imaging features, is subject to significant subjective variation and only captures partial morphological features of the tumor, failing to fully reflect its biological behavior. While deep learning has been applied in this field to extract features from images, it relies heavily on large amounts of high-quality data, making it difficult to ensure accurate and consistent data annotation and model interpretability. Furthermore, existing prediction methods are often based on a single data source, making it difficult to integrate multidimensional information and lacking effective integration between different methods. For example, traditional clinical indicator predictions fail to consider imaging features, while imaging predictions ignore clinicopathological information. Deep learning models are insufficiently integrated with clinical data.

[0029] Therefore, a new prediction method is urgently needed to improve prediction accuracy.

[0030] In view of this, an embodiment of the present application provides a method for predicting the postoperative progression of lung adenocarcinoma. First, a clinical data set and CT imaging data of postoperative lung adenocarcinoma patients are obtained. The clinical data covers multiple clinical parameters and corresponding postoperative condition labels. Then, a clinical data model is constructed based on the collected clinical data set. By analyzing the model, parameters with significant predictive effects on the clinical data model are screened out from multiple clinical parameters. These parameters are independent predictors. Next, an imaging genomics feature model is constructed for the CT imaging data, and feature weights are obtained based on the imaging genomics feature model. Subsequently, an image feature extraction model is constructed using the CT imaging data and the pre-trained model. Through this model, a learning feature score reflecting the postoperative characteristics of lung adenocarcinoma can be further obtained. After obtaining the independent predictors, feature weights, and learning feature scores, a target prediction model is constructed based on these data. In actual application, the postoperative lung adenocarcinoma data to be predicted is input into the target prediction model to obtain the prediction results of the postoperative progression of lung adenocarcinoma. The embodiment of the present application integrates independent predictive factors obtained based on a clinical data model, feature weights obtained based on an imaging genomics feature model, and learning feature scores obtained based on an image feature extraction model to build a model, surpassing the performance of a single data source model, thereby effectively improving the accuracy of postoperative prediction of lung adenocarcinoma.

[0031] The embodiments of the present application are further described below with reference to the accompanying drawings.

[0032] Reference Figure 1 , Figure 1 This is a flow chart of a method for predicting lung adenocarcinoma progression after surgery provided in an embodiment of the present application. The method includes but is not limited to steps 110 to 160.

[0033] Step 110: Acquire a clinical dataset and CT image data of lung adenocarcinoma after surgery, wherein the clinical dataset includes multiple clinical parameters and postoperative condition labels;

[0034] Step 120: constructing a clinical data model based on the clinical data set, and obtaining independent predictors based on the clinical data model, wherein the independent predictors are parameters among the multiple clinical parameters that have a predictive effect on the dependent variable of the clinical data model;

[0035] Step 130: constructing a radiomics feature model based on the CT image data, and obtaining feature weights based on the radiomics feature model;

[0036] Step 140: constructing an image feature extraction model based on the CT image data and the pre-trained model, and obtaining a learning feature score based on the image feature extraction model;

[0037] Step 150: constructing a target prediction model based on the independent predictor, feature weights, and learned feature scores;

[0038] Step 160: Input the postoperative lung adenocarcinoma data to be predicted into the target prediction model to obtain a prediction result.

[0039] Steps 110 to 160 are described in detail below.

[0040] In one feasible embodiment, a postoperative clinical dataset for lung adenocarcinoma is a collection of various information related to patients after surgical treatment, including multiple clinical parameters and postoperative status labels. These multiple clinical parameters cover basic patient characteristics, disease-related characteristics, and tumor marker indicators. Basic characteristics such as gender, age, and smoking history are associated with the development and progression of lung adenocarcinoma. Disease-related characteristics such as tumor location, average size, borders, morphology, lobulation, spiculation, vacuolation, air bronchograms, emphysema, and stage can intuitively reflect tumor characteristics and disease severity. Tumor marker indicators such as Cyfra211, NSE, CEA, and CA199, their numerical changes, can help determine tumor activity and treatment efficacy. Postoperative status labels record whether the patient has progressed during postoperative recovery, such as whether the tumor has recurred or metastasized. These labels can be explicitly set, for example, "0" indicates no postoperative progression (i.e., no recurrence or metastasis) and "1" indicates postoperative progression (i.e., recurrence or metastasis). CT images can provide detailed image information of lung tumors and surrounding tissues. By analyzing CT images, we can observe the tumor's morphology, size, location, density, and relationship with surrounding tissues. These imaging features help identify potential lesions, supplement the deficiencies of clinical data sets, and provide a basis for more accurate disease assessment and prediction of postoperative progression.

[0041] In one feasible embodiment, before building a clinical data model based on a clinical dataset, the clinical dataset can be preprocessed to improve the quality and reliability of model construction. Specifically, when performing preprocessing analysis on the clinical dataset, descriptive statistical analysis should be performed first, and the data should be grouped according to whether there is postoperative progression. Descriptive statistical analysis can provide a preliminary understanding of the basic characteristics of the clinical dataset, such as the mean, median, and standard deviation of each clinical parameter. After completing the descriptive statistical analysis and grouping, the clinical parameters can be further subjected to correlation analysis, normality test, and homogeneity of variance test. These test steps ensure that the data meets the assumptions for subsequent model construction, thereby improving the accuracy and reliability of the model. For correlation analysis, the Spearman correlation coefficient significance test can be used to calculate the correlation coefficients between the clinical parameters and plot a visualization of the correlation coefficient matrix. This visualization allows intuitive observation of the correlations between clinical parameters and the presence of collinearity. Specifically, if the correlation coefficient between two clinical parameters is close to 1, it indicates a strong positive correlation; if it is close to 1, it indicates a strong negative correlation. During the subsequent construction of clinical data models, strongly correlated clinical parameters are likely to cause collinearity. Collinearity can lead to model instability, significantly reducing the accuracy of parameter estimates and complicating model interpretation. Therefore, when collinearity is detected, appropriate measures can be taken, such as removing clinical parameters with excessive correlation or performing dimensionality reduction using methods such as principal component analysis. Clinical parameters identified as free of collinearity through correlation analysis will serve as valid data for constructing the clinical data model. The independence of these data can avoid model bias caused by collinearity, enabling the constructed model to more accurately reflect the true relationship between clinical parameters and postoperative progression, improving the model's predictive accuracy and stability. For normality testing, the Shapiro-Wilk normality test can be used to determine whether the data conform to a normal distribution. Failure to meet the normal distribution assumption can lead to biased statistical test results. For example, using statistical methods based on the normal distribution assumption on non-normally distributed data may lead to erroneous conclusions. Therefore, when data are found to not conform to a normal distribution, data transformations, such as logarithmic or square root transformations, can be considered to bring them as close to a normal distribution as possible to meet the requirements of subsequent analysis. When building clinical data models, many commonly used statistical methods and machine learning algorithms, such as linear regression and logistic regression, require data normality. Data processed for normality testing can ensure more accurate and reliable parameter estimates for these models, improving the models' predictive performance and explanatory power.For example, in a logistic regression model, data that adheres to a normal distribution can make the model coefficient estimates more stable, thereby more accurately predicting the probability of postoperative progression of lung adenocarcinoma. Regarding homogeneity of variance testing, the Bartlett test can be used to test whether the variances are equal across groups. Homogeneity of variance is a critical prerequisite for many statistical analysis methods. If the variances are unequal, the accuracy and reliability of the analysis results will be affected. For example, when performing methods such as analysis of variance, failure to meet the assumption of homogeneity of variance can lead to reduced test power and even incorrect conclusions. When unequal variances are found between groups, nonparametric tests can be considered, or appropriate data transformations can be performed to ensure homogeneity of variance. Homogeneity of variance is also crucial for ensuring model stability and accuracy when constructing clinical data models. For example, when performing group comparisons and analyses, homogeneity of variance ensures that differences between groups are real and not misjudgments due to inconsistent variances. Data processed through the homogeneity of variance test can make the constructed clinical data model more robust, improving the model's generalization ability and predictive accuracy across different groups. Through the above-mentioned correlation analysis, normality test and homogeneity of variance test, problems in the data can be discovered in advance, and corresponding treatment measures can be taken according to the specific situation, such as data conversion, clinical parameter screening, etc., to effectively ensure that the data meets the assumptions of subsequent model construction.

[0042] In a feasible embodiment, after performing correlation analysis, normality test and variance homogeneity test on these clinical parameters, a clinical data model can be further constructed based on the clinical data set, and independent predictive factors can be obtained based on the clinical data model. Figure 2 As shown, the specific process of step 120 may at least include steps 210 to 240.

[0043] Step 210: Modeling is performed based on each clinical parameter and postoperative condition label in the clinical data set to obtain a logistic regression model for each clinical parameter;

[0044] Step 220: Determine the relationship between the postoperative condition label and each clinical parameter based on the logistic regression model of each clinical parameter;

[0045] Step 230: Filter out clinical parameters whose relationship values ​​are less than a preset threshold from the clinical data set to obtain a set of predictive factors;

[0046] Step 240: Identify independent predictors from the set of predictors.

[0047] In a feasible embodiment, step 210 aims to establish a quantitative relationship model between each clinical parameter and the postoperative condition label. Through the statistical method of logistic regression, it is possible to clarify the influence of each clinical parameter on the result of whether the patient progresses after surgery. Specifically, each clinical parameter in the clinical data set can be used as an independent variable, such as the patient's age, gender, tumor size, pathological stage, etc.; and the postoperative condition label can be used as the dependent variable. The data are processed using a logistic regression algorithm to obtain a logistic regression model that reflects the relationship between each clinical parameter and postoperative progression. For example, for the clinical parameter "age", the logistic regression model can calculate the change in the probability of postoperative progression as age changes.

[0048] In a feasible embodiment, step 220 is intended to extract quantitative indicators that can accurately measure the degree of association between the postoperative condition label and each clinical parameter from the constructed logistic regression model of each clinical parameter, so as to subsequently screen out clinical parameters that are significantly correlated with postoperative progression. Specifically, key indicators such as coefficients, confidence intervals, and p-values ​​can be extracted from the logistic regression model of each clinical parameter, wherein the coefficients can characterize the direction and intensity of the influence of the clinical parameters on the postoperative condition label. For example, a positive coefficient indicates that as the value of the clinical parameter increases, the probability of postoperative progression (labeled as 1) increases; a negative coefficient indicates the opposite. The confidence interval is used to measure the reliability of the coefficient estimate and reflects the range of values ​​of the coefficient at a certain confidence level. The p-value, as a relationship value, can be used to test whether there is a statistically significant relationship between the clinical parameter and the postoperative condition label. Generally, the smaller the p-value, the greater the possibility of a significant relationship between the two.

[0049] In a feasible embodiment, step 230 is to screen out clinical parameters that are significantly associated with postoperative progression based on the relationship value obtained in step 220 by setting a reasonable preset threshold. The set of these parameters can be used as a candidate set for subsequent determination of independent predictive factors. Specifically, a preset threshold is set, for example, the preset threshold is 0.05. The p-value of each clinical parameter obtained in step 220 is compared with the preset threshold, and clinical parameters with a p-value less than the preset threshold are screened out. These screened clinical parameters have a significant relationship with the postoperative condition label, and they are combined together to form a predictor set. For example, among all clinical parameters, if the p-value of parameters such as "tumor size", "pathological stage" and "smoking history" is less than 0.05, these parameters are included in the predictor set because they have a statistically significant impact on whether postoperative progression occurs and can be used as key factors for predicting postoperative progression.

[0050] In a feasible embodiment, the purpose of step 240 is to further screen out those factors that have independent predictive effects on postoperative progression from the set of predictive factors, that is, these factors are not interfered with by other factors and can independently predict postoperative progression, thereby providing a key basis for constructing a clinical data model.

[0051] See also Figure 3 The specific process of confirming the independent predictor from the predictor set in step 240 may include but is not limited to steps 310 to 360.

[0052] Step 310: Constructing a first clinical data model based on the postoperative condition labels;

[0053] Step 320: adding the predictor set to the first clinical data model, and calculating a first significance relationship value between the predictor set and the postoperative condition label;

[0054] Step 330: Filter out the prediction factor set corresponding to the first meaning relationship value being less than a preset threshold to obtain a first initial independent prediction factor;

[0055] Step 340: Modeling based on the predictor set and the postoperative condition label to obtain a second clinical data model;

[0056] Step 350: Calculate the second significance relationship value corresponding to each prediction factor set in the second clinical data model, and filter out the prediction factor sets whose second significance relationship values ​​are less than a preset threshold value, until the second significance relationship values ​​corresponding to each prediction factor set in the second clinical data model are all less than the preset threshold value, thereby obtaining the second initial independent prediction factors;

[0057] Step 360: Confirm the independent predictor based on the first initial independent predictor and the second initial independent predictor.

[0058] In a feasible embodiment, step 310 is intended to construct a basic model (i.e., a first clinical data model) based on the postoperative condition label to provide an initial model framework for subsequent analysis of the relationship between the predictor and the postoperative condition. Specifically, a suitable statistical or machine learning method can be selected to construct the first clinical data model. For example, a logistic regression method can be used with the postoperative condition label as the dependent variable. It should be noted that in the logistic regression model, the postoperative condition label is used as the output variable. Although the predictor set is not considered at this time, the model can be preliminarily modeled based on the distribution and characteristics of the postoperative condition label itself to prepare for the subsequent addition of predictors.

[0059] In a feasible embodiment, step 320 introduces a set of predictors based on the first clinical data model, aiming to analyze the degree of association between the predictors and the postoperative condition labels. By calculating the first significance relationship value, the influence and significance of each predictor on the postoperative condition label can be quantified, providing a basis for the subsequent screening of meaningful predictors. Specifically, each predictor in the predictor set can be added to the first clinical data model as an independent variable. In the case where the first clinical data model is a logistic regression model, the model becomes a multivariate logistic regression model containing multiple independent variables (predictors) and one dependent variable (postoperative condition label). In the process of calculating the first significance relationship value, the p-value corresponding to each predictor can be extracted from the multivariate logistic regression model, and this p-value is the first significance relationship value between the predictor set and the postoperative condition label. The smaller the p-value, the greater the possibility that there is a significant association between the predictor and the postoperative condition label.

[0060] In a feasible embodiment, in step 330, based on the first significance relationship value calculated in step 320, a preset threshold value can be used for screening, and those predictors that are not significantly related to the postoperative condition label can be removed, thereby obtaining a set of predictors with a certain significance, namely the first initial independent predictor. These factors show a close relationship with the postoperative progress in the preliminary screening, which narrows the scope for further determination of independent predictors. Specifically, the first significance relationship value of each predictor is compared with the preset threshold value (0.05), and predictors with a p value less than the preset threshold value are screened out. These screened predictors constitute the first initial independent predictor set, which has a significant impact on the postoperative condition label in the current model.

[0061] In one feasible embodiment, the second clinical data model constructed in step 340, based on the predictor set and the postoperative condition label, is intended to comprehensively evaluate the relationship between the predictors and the postoperative condition label from another perspective, providing more information and verification for further screening of independent predictors. Specifically, a logistic regression method can also be selected, constructing the model using each predictor in the predictor set as an independent variable and the postoperative condition label as the dependent variable. During the construction process, the data characteristics of the predictor set and the postoperative condition label are fully utilized to enable the model to more accurately reflect the relationship between the two.

[0062] In a feasible embodiment, the purpose of step 350 is to further evaluate the significance of each predictor by calculating the second significance relationship value based on the second clinical data model. Continuously screen out predictors with significant significance until all predictors in the model meet the preset significance criteria, thereby obtaining a second initial independent predictor. This step is a further screening and optimization of the predictors to ensure that the factors finally obtained can more accurately predict postoperative progression. Specifically, the second significance relationship value (p value) corresponding to each predictor can be extracted from the second clinical data model. These values ​​are compared with a preset threshold (0.05) to screen out predictors with p values ​​less than the preset threshold. If there are still predictors with p values ​​greater than the preset threshold in the model after screening, these insignificant factors are removed from the model, the model is rebuilt and the second significance relationship value is calculated again, and this process is repeated until the p values ​​of all predictors in the model are less than the preset threshold, obtaining a second initial independent predictor set composed of these predictors.

[0063] In a feasible embodiment, in step 360, the information of the first initial independent predictor and the second initial independent predictor is integrated to perform a final independent predictor confirmation. Since the two initial independent predictor sets are obtained through different models and screening methods, comprehensive consideration can improve the accuracy and reliability of the independent predictor, ensuring that the independent predictor finally determined can truly independently and effectively predict the postoperative progression. Specifically, the first initial independent predictor set and the second initial independent predictor set are compared and analyzed. The intersection of the two sets can be taken, that is, the predictors that exist in both sets at the same time. These factors show significance under two different screening methods and have high reliability, and can be directly confirmed as independent predictors. In addition, the factors in the two sets can be further combined with domain knowledge and actual clinical experience to conduct a comprehensive evaluation and adjustment of the factors in the two sets, and finally determine the independent predictor set that can most accurately predict the postoperative progression.

[0064] In a possible embodiment, if Figure 4 As shown, the specific process of constructing the radiomics feature model based on the CT image data in step 130 may include but is not limited to steps 410 to 440.

[0065] Step 410: marking the lesion area on each CT image in the CT image data to obtain a key area image corresponding to each CT image;

[0066] Step 420: extract features from the key area image to obtain an image feature dataset;

[0067] Step 430: performing feature screening processing on the image feature dataset to obtain a modeling feature dataset;

[0068] Step 440: Construct a radiomics feature model based on the modeling feature dataset.

[0069] In a feasible embodiment, in step 410, each CT image in the CT image data is annotated with lesion areas, with the purpose of accurately locating the lesion areas related to lung adenocarcinoma in the CT image, removing irrelevant background information in the entire CT image, and focusing on the key areas. By annotating the lesion areas, it is possible to highlight the possible presence of tumor tissue, which facilitates the subsequent in-depth analysis of the characteristics of these areas, thereby better understanding the imaging characteristics of lung adenocarcinoma. Generally, the CT image data exported from the medical center usually contains multiple CT images, and the original image size of each image is 512×512×n (n represents the number of layers of the image), such as Figure 5 As shown. In the actual operation process, when marking the lesion area of ​​the CT image, each CT image can be carefully observed and analyzed under the mediastinal window (window width is 350HU, window level is 50HU). The setting of the mediastinal window helps to show the tissue structure and lesion condition of the lung more clearly. Then use professional image annotation tools to accurately outline the lesion area. The outlined area is the key area, which is the key area for subsequent feature extraction. In this way, the key area images corresponding to each CT image are obtained, as shown Figure 6 shown.

[0070] In one feasible embodiment, step 420 aims to extract various information reflecting the imaging characteristics of lung adenocarcinoma from the key region images to form an image feature dataset. These features can describe the lesion's morphology, texture, density, and other information from different perspectives, providing rich data dimensions for subsequent model construction, enabling a more comprehensive analysis of the relationship between the imaging characteristics of lung adenocarcinoma and postoperative progression. A specific operation may include: first normalizing the key region images with the lesion area annotated. Normalizing the CT image data can uniformly scale and distribute the data, facilitating subsequent calculations and analysis. Simultaneously, the images are resampled, adjusting the pixel spacing to 1.5 mm × 1.5 mm × 1.5 mm. This helps to standardize the spatial resolution of the images and reduce variations caused by different scanning devices or parameter settings. The images are then processed using three Gaussian filters with different standard deviations. Gaussian filters can smooth the images, reduce the impact of noise, and preserve the key features of the images. Histogram equalization uses a bin width of 5 per bin to adjust the image contrast and enhance detail clarity. After the above preprocessing, the pyradiomics library can be used to extract radiomic features from images with key regions. This library extracts first-order features (such as mean, variance, and skewness, which describe basic statistical characteristics of the image), second-order features (such as gray-level co-occurrence matrix-related features, which reflect the spatial relationships between pixels in the image), and higher-order features, resulting in multiple (for example, 2004) radiomic features. These features are combined to form an image feature dataset.

[0071] In one feasible embodiment, since the image feature dataset obtained in step 420 contains a large number of features, some of which may contribute little to the model, or may contain redundancy and noise, and may even interfere with model performance. Therefore, the purpose of step 430 is to remove unimportant or irrelevant features through a series of feature screening methods, while retaining features that are significantly effective in predicting postoperative progression of lung adenocarcinoma, thereby obtaining a modeling feature dataset and improving the efficiency and accuracy of subsequent model construction.

[0072] In one feasible embodiment, during feature screening of an image feature dataset to obtain a modeling feature dataset, the image feature dataset can be first screened to obtain a preliminary feature dataset. A one-way analysis of variance (ANOVA) is then performed on the preliminary feature dataset to obtain an intermediate feature dataset. Regression analysis is then performed on the intermediate feature dataset, and the modeling feature dataset is screened based on the analysis results. Specifically, the image feature dataset is first preprocessed to remove non-numeric features, as subsequent analysis and modeling are typically based on numerical data. A variance analysis is then used to remove features with low divergence, such as features with a variance less than 0.001. Features with low variance indicate that their values ​​are relatively concentrated. Retaining such features not only contributes little to the model but may also increase the computational load and model complexity. Therefore, these features are removed to obtain a preliminary feature dataset. A one-way analysis of variance (ANOVA) is then performed to determine whether each feature is statistically significant by calculating the variance difference between different categories (e.g., those with postoperative progression and those without progression). Specifically, when the p-value corresponding to a feature is less than 0.05, it indicates that the feature significantly differs between different categories, indicating that the feature has a certain role in distinguishing postoperative progression. Therefore, features with p-values ​​less than 0.05 are selected to form the intermediate processing feature dataset, as these features are more likely to be associated with postoperative progression. To further optimize the feature dataset, during the process of performing regression analysis on the intermediate processing feature dataset and filtering the modeling feature dataset based on the analysis results, a ten-fold cross-validation can be performed to select the optimal λ value for lasso feature screening. Lasso (least absolute shrinkage and selection operator) is a feature selection and regularization method that controls the number of selected features and model complexity by adjusting the λ value. In the ten-fold cross-validation process, the dataset is evenly divided into ten parts, with nine parts used for model training and one part used for validation. This is repeated ten times to evaluate model performance under different λ values. The λ value that optimizes model performance is selected and then used for lasso feature screening, resulting in a more representative and effective feature subset. During the feature screening process, a random feature dropout strategy can be implemented, randomly dropping some features during each training step. This strategy can effectively prevent overfitting of the constructed radiomics feature model and improve the model's generalization ability. Overfitting refers to the phenomenon where a model performs well on training data but poorly on unseen test data. By randomly discarding features, the model can learn more robust feature patterns, reduce its reliance on specific features, and thus have better predictive ability when faced with new data. After the above series of screening steps, the modeling feature dataset was finally obtained. This dataset contains features that have a significant impact on the postoperative progression of lung adenocarcinoma.

[0073] In a feasible embodiment, step 440 is intended to use the modeling feature data set screened in step 430 to construct an imaging omics feature model that can accurately describe the relationship between the imaging features of lung adenocarcinoma and postoperative progression. This model can be used to predict the postoperative progression of patients with lung adenocarcinoma, providing an important reference for clinical diagnosis and treatment. The specific operation method includes: using the screened modeling feature data set as input, selecting a suitable modeling method, such as logistic regression, etc. Each feature in the modeling feature data set is used as an independent variable, and the postoperative condition label (such as "0" for no progression and "1" for progression) is used as a dependent variable to construct an imaging omics feature model. By training and optimizing the model, the parameters of the model are adjusted so that the model can fit the data as accurately as possible, thereby obtaining an effective imaging omics feature model.

[0074] In a feasible embodiment, after obtaining the imaging omics feature model, the feature weight can be obtained based on the imaging omics feature model. The feature weight can determine which features play a key role in the prediction of postoperative progression of lung adenocarcinoma and which features have relatively small effects. Specifically, the regression coefficient of the modeling feature data set can be extracted from the imaging omics feature model. It should be noted that in the logistic regression model, the regression coefficient reflects the direction and degree of influence of each feature on the dependent variable (postoperative condition label). Then, the feature weight of each feature is obtained based on the absolute value of the regression coefficient. The larger the absolute value of the regression coefficient, the greater the influence of the feature on the prediction result, and the higher its feature weight. In this way, the importance of each feature in the model can be quantified.

[0075] In a possible embodiment, if Figure 7 As shown, the specific process of constructing the image feature extraction model based on the CT image data and the pre-trained model in step 140 may include but is not limited to steps 710 to 740.

[0076] Step 710: resampling, lesion area marking, and stretching processing are performed on all CT images in the CT image data to obtain target processed images corresponding to each CT image;

[0077] Step 720: Adjust parameters of the pre-trained model based on the CT image data to obtain a target feature extraction model;

[0078] Step 730: Using the target feature extraction model, perform feature extraction and feature screening on the target processed images corresponding to each CT image to obtain target image features;

[0079] Step 740: Construct an image feature extraction model based on the target image features.

[0080] In a feasible embodiment, the purpose of step 710 is to pre-process the original CT image data so that it meets the requirements of subsequent model training and feature extraction. Through resampling, lesion area annotation and stretching processing, the resolution and size of the CT image can be unified, the lesion area can be highlighted, and high-quality data can be provided for building an accurate image feature extraction model. The specific operation method includes: first, each original CT image in the CT image data (the original image size is 512×512×n) is resampled and its resolution is adjusted to 1mm×1mm×1mm. Resampling can make CT images obtained with different scanning parameters have a uniform spatial resolution and reduce data differences. Next, the lesion area in the CT image is annotated. For example, the lesion area in the CT image can be accurately annotated under the mediastinal window (window width is 350HU, window level is 50HU) to determine the region of interest. This step is to accurately locate the area related to the disease and exclude irrelevant background information. Finally, the CT images were cropped according to the region of interest and stretched to obtain a 48 mm × 48 mm × 48 mm processed image containing the lesion area and its surroundings. This process ensures that each CT image has the same size and format, facilitating subsequent processing and analysis.

[0081] In one feasible embodiment, in step 720, a model that has been pre-trained on a large-scale dataset is used in conjunction with the current CT image data. By adjusting the model's parameters to adapt it to extract features related to lung adenocarcinoma from CT images, a target feature extraction model specifically tailored to this task can be obtained. This allows the general feature representations learned by the pre-trained model to be utilized, reducing model training time and data requirements, and improving model performance. Specifically, when constructing the target feature extraction model, a suitable pre-trained model can be selected and transfer learning can be implemented according to the following process: First, an adapted pre-trained model, such as FcaNet152, is selected. This model has been trained on other related tasks or large-scale datasets and possesses certain general feature extraction capabilities. Subsequently, the CT image data is pre-processed to ensure that the data format and size meet the input requirements of the pre-trained model. The pre-processed CT image data is then input into the pre-trained model. The model then processes the input data based on its learned feature patterns and outputs a result. Next, the difference between the model output and the actual label (e.g., whether the patient progressed after surgery) is calculated. This difference can be measured using methods such as cross-entropy loss and mean squared error. Based on the calculated differences, the parameters of the pre-trained model are adjusted using the transfer learning method. Transfer learning aims to transfer the knowledge learned by the pre-trained model in the previous task to the new lung adenocarcinoma feature extraction task. By fine-tuning the model parameters, the model is adapted to the new CT image dataset and lung adenocarcinoma feature extraction task. During the training process, the dataset is usually divided into a training set, a validation set, and a test set. The training set is used to adjust the model parameters, and the validation set is used to evaluate the performance of the model during the training process to prevent the model from overfitting. Through continuous iterative training, the model parameters are continuously adjusted until the performance indicators of the model on the validation set (such as accuracy, recall rate, F1 value, etc.) tend to be stable and meet the expected standards. After a series of training and parameter adjustments, the target feature extraction model can be obtained.

[0082] In a feasible embodiment, step 730 aims to extract rich features from the target processed image through the target feature extraction model, and after a series of feature screening processes, remove unimportant or irrelevant features, retain features that have a significant effect on distinguishing between diseased types and non-disease types, and obtain target image features, thereby providing key input data for the subsequent construction of the image feature extraction model.

[0083] See also Figure 8 In step 730, the target feature extraction model is used to perform feature extraction and feature screening on the target processing images corresponding to each CT image to obtain the target image features. The specific execution process may include but is not limited to steps 810 to 860.

[0084] Step 810: Merging target processed images corresponding to the CT images to obtain multiple target image blocks, each target image block including a first processed image, a second processed image, and a third processed image, wherein the first processed image corresponds to the first color channel, the second processed image corresponds to the second color channel, and the third processed image corresponds to the third color channel;

[0085] Step 820: Using the target feature extraction model, extract features from the target image block to obtain preliminary image feature data;

[0086] Step 830: Divide the preliminary image feature data into channels according to the first color channel, the second color channel, and the third color channel to obtain first channel feature data, second channel feature data, and third channel feature data;

[0087] Step 840: extracting a first target feature from the first channel feature data, extracting a second target feature from the second channel feature data, and extracting a third target feature from the third channel feature data to obtain target channel feature data;

[0088] Step 850: performing feature screening processing on each feature in the target channel feature data to obtain intermediate image feature data;

[0089] Step 860: Using ten-fold cross validation, perform feature screening on the intermediate image feature data to obtain target image features.

[0090] In a feasible embodiment, before merging the target processed images corresponding to the CT images, these target processed images can be grayscale processed to obtain target grayscale processed images corresponding to the target processed images, and then these target grayscale processed images are merged. The specific operation process includes: combining every three grayscale images (i.e., the first processed image, the second processed image, and the third processed image) into an RGB image, thereby forming multiple target image blocks. The first processed image in each target image block corresponds to the red channel (first color channel) of the RGB image, the second processed image corresponds to the green channel (second color channel), and the third processed image corresponds to the blue channel (third color channel). The size of the final RGB image is 48mm×48mm×3 (3 represents three color channels). In this way, multiple target processed images can be combined into multiple target image blocks, each of which contains information from three different color channels.

[0091] In a feasible embodiment, in step 820, the target image block obtained in step 810 is input into the target feature extraction model, and the model can mine valuable feature information from the merged target image block. These features can reflect the texture, shape, structure and other aspects of the CT image related to lung adenocarcinoma. Specifically, the target feature extraction model can use its pre-trained weights and network structure to perform operations such as convolution and pooling on the target image block to extract features from various levels and scales of the image. Through these operations, the model can automatically learn the complex feature patterns in the image and ultimately output preliminary image feature data, which contains rich image feature information, for example, may include 94,208 features.

[0092] In a feasible embodiment, in order to more carefully analyze and process the feature information contained in different color channels, the preliminary image feature data can be divided according to the color channel. Different color channels may be sensitive to different image features. Processing the feature data of each channel separately helps to dig out more targeted features and improve the distinguishability and representativeness of the features. Specifically, the preliminary image feature data can be split according to the color channel information corresponding to each processed image in the target image block. The features related to the first color channel (red channel) are extracted to form the first channel feature data; similarly, the features related to the second color channel (green channel) and the third color channel (blue channel) are extracted separately to form the second channel feature data and the third channel feature data.

[0093] In one feasible embodiment, in step 840, the most representative and discriminative features are further screened and extracted from the feature data of each channel. This can remove redundant and irrelevant features, making subsequent analysis and modeling more efficient and accurate. Specifically, the maximum value of each layer of the image is taken for each channel. A representative feature is selected from the first channel feature data as the first target feature, a corresponding second target feature is selected from the second channel feature data, and a third target feature is selected from the third channel feature data. After this processing, the target channel feature data is ultimately obtained.

[0094] In a feasible embodiment, in step 850, the target channel feature data is screened by statistical methods, and those features that are not significantly correlated with disease diagnosis and prediction can be removed, and features with statistical significance can be retained. In this way, the feature set can be further optimized, the generalization ability and prediction accuracy of the model can be improved, and overfitting of the model due to the inclusion of too many irrelevant features can be avoided. Specifically, the single-factor analysis of variance method can be used to analyze each feature in the target channel feature data. The variance difference between different categories (such as lung adenocarcinoma with progression and no progression after surgery) of each feature is calculated, and the corresponding p-value is obtained. Features with a p-value less than 0.05 are selected. These features have significant differences between different categories, indicating that they have a certain effect on distinguishing disease states. These statistically significant features are retained to form the intermediate image feature data.

[0095] In a feasible embodiment, in step 860, in order to further optimize the feature set and improve the stability and generalization ability of the model, a ten-fold cross-validation combined with feature screening method can be used. Ten-fold cross-validation can fully utilize the data to evaluate the performance of the model under different data set partitions, avoiding the deviation of the model evaluation results due to the randomness of the data partitioning. Through screening, the feature subset that best suits the model can be found, thereby improving the model's predictive performance. In actual operation, the intermediate image feature data can be evenly divided into ten parts, and nine of them are selected as training sets and one as validation sets each time. During each training process, the l asso feature screening method is used to control the number of selected features and the complexity of the model by adjusting the lambda value. At the same time, a random feature dropout strategy is added to prevent model overfitting. During the ten-fold cross-validation process, the performance of the model under different feature subsets is continuously evaluated, and the feature subset that optimizes the model performance is selected as the final target image feature. After such processing, a set of features with the most representative and predictive power can be obtained for subsequent logistic regression modeling and disease prediction.

[0096] In a feasible embodiment, the purpose of step 740 is to use the target image features obtained in step 730 to construct an image feature extraction model that can accurately extract features related to lung adenocarcinoma from CT images. Specifically, the target image features can be used as input, and the postoperative progression of lung adenocarcinoma patients (such as "0" for no progression and "1" for progression) can be used as output to construct a logistic regression model. The logistic regression model can estimate the parameters of the model through methods such as maximum likelihood estimation, so that the model can fit the training data as accurately as possible. During the training process, the parameters of the model are continuously adjusted to minimize the error between the prediction results of the model on the training set and the true label, so that the model can accurately capture the feature information related to lung adenocarcinoma in the CT image, and then construct an image feature extraction model.

[0097] In one feasible embodiment, a learning feature score (also known as a deep learning feature score) can be used to quantify the importance and effectiveness of features extracted by the image feature extraction model for disease prediction. This calculation can be based on the target image features obtained after screening and is performed using the image feature extraction model. Specifically, in the process of obtaining a deep learning feature score based on the image feature extraction model, the regression coefficient set output by the image feature extraction model can be first obtained from the image feature extraction model. This set is a summary of the regression coefficients determined for each feature in the target image during model training, based on the correlation between the target image features and the postoperative condition labels. These regression coefficients reflect the direction and degree of influence of each feature on the prediction results. Furthermore, from this regression coefficient set, the regression coefficient corresponding to each specific feature in the target image can be matched and determined. Based on the regression coefficient corresponding to each feature, a deep learning feature score for the target image feature is obtained through a specific calculation method. The specific calculation process may include weighting each feature according to the regression coefficient derived by the model. For each sample (patient's CT image), the feature value of each feature is multiplied by the corresponding regression coefficient, and these products are then summed. The resulting composite score is the learning feature score for that sample. The higher the learning feature score, the more critical the feature combination corresponding to the sample plays in disease prediction, and the higher its value for disease diagnosis and prediction.

[0098] In a possible embodiment, if Figure 9 As shown, step 150 constructs a specific process of obtaining a target prediction model based on independent prediction factors, feature weights, and learning feature scores, which may include but is not limited to steps 910 to 940.

[0099] Step 910: Standardize the independent predictors and the learning feature scores, and construct an initial prediction model based on the processed independent predictors, learning feature scores, and feature weights;

[0100] Step 920: Obtain a prediction training set and a prediction test set;

[0101] Step 930: Train the initial prediction model based on the prediction training set, and adjust the weight parameters of the initial prediction model based on a preset error index. When the maximum number of iterations is reached or the error converges to a preset threshold, the initial prediction model is obtained.

[0102] Step 940: Evaluate the initial prediction model using the prediction test set. When the performance of the initial prediction model on the prediction test set meets the requirements, use the initial prediction model as the target prediction model.

[0103] Understandably, independent predictors and learned feature scores may have different dimensions and value ranges. Using them directly can lead to unstable model training and difficulty in achieving accurate convergence. Standardization can eliminate these differences, aligning all data with the same scale, improving model training accuracy and generalization. Furthermore, combining feature weights can highlight the importance of different factors in the model, thereby constructing an initial prediction model that comprehensively considers these factors.

[0104] In a feasible embodiment, when standardizing the independent predictors and learning feature scores, an appropriate standardization method, such as the common Z-score standardization, can be used to convert the independent predictors and learning feature scores to a unified scale to ensure that the model treats each factor equally during the training process. Furthermore, the standardized independent predictors and learning feature scores are combined with the corresponding feature weights. Assuming that the model is constructed using a linear combination method, it can be expressed as y = w1x1 + w2x2 + ... + w n x n , where x i (i=1,2,...,n) is the independent predictor or learning feature score after treatment, w i is the corresponding feature weight, and y is the prediction result.

[0105] In a feasible embodiment, in order to train and evaluate the initial prediction model, the data needs to be divided into a training set and a test set. The training set is used to allow the model to learn the patterns and regularities in the data and adjust the model parameters; the test set is used to evaluate the performance of the model on unseen data, verify the generalization ability of the model, and ensure that the model is reliable in practical applications. For example, its operation method may include: randomly dividing the prediction training set and the prediction test set from the original data set according to a certain proportion (such as the commonly used 70%-30% or 80%-20%). In order to avoid the problem of the model being biased towards a certain category due to unbalanced data division, a stratified sampling method can be used to ensure that the proportion of each category (such as postoperative recurrence and non-recurrence) in the training set and the test set is consistent with the original data set. In this way, the model can learn and evaluate different situations more comprehensively during the training and testing process, thereby improving the applicability of the model.

[0106] In one feasible embodiment, an initial prediction model is trained using a prediction training set. By continuously adjusting weight parameters, the model is better fitted to the training data, reducing the error between the predicted results and the true labels. Setting a maximum number of iterations and an error convergence threshold is intended to control the training process, prevent the model from overfitting or falling into an infinite loop, and ensure that the model achieves good performance within a reasonable time and computing resources. For example, the operation may include: inputting data from the prediction training set into the initial prediction model, the model making predictions based on the current weight parameters, and calculating the error between the predicted results and the true labels. Common error metrics include mean squared error (MSE) and cross entropy loss. By calculating the error, the degree of deviation between the model prediction and the true situation can be measured. Furthermore, based on the calculated error, an optimization algorithm (such as gradient descent) can be used to adjust the weight parameters of the initial prediction model. Gradient descent calculates the gradient of the error with respect to the weight parameters and updates the weight parameters in the opposite direction of the gradient, gradually reducing the error. During the training process, it is continuously checked whether the maximum number of iterations has been reached or whether the error has converged to a preset threshold. If the maximum number of iterations is reached, it indicates that sufficient rounds of training have been performed. If the error converges to the preset threshold, it indicates that the model performance has stabilized. At this point, training is stopped and the trained initial prediction model is obtained.

[0107] In a feasible embodiment, the trained initial prediction model is evaluated using a prediction test set to test the performance of the model on unseen data. Only when the performance of the model on the test set meets the preset requirements can it be used as the final target prediction model to ensure that the model has good generalization ability and reliability and can be applied to actual disease prediction scenarios. For example, its operation method may include: inputting the data in the prediction test set into the initial prediction model to obtain the prediction results. Then, a series of evaluation indicators (such as AUC, accuracy, sensitivity, specificity, PPV and NPV, etc.) are used to evaluate the performance of the model. The AUC (area under the curve) measures the model's overall discriminatory ability, ranging from 0.5 to 1. Values ​​closer to 1 indicate a stronger model's ability to distinguish between positive and negative examples. Accuracy is the ratio of correctly predicted samples to the total number of samples. Sensitivity (true positive rate) is the ratio of correctly predicted positive samples to the total number of positive samples. Specificity (true negative rate) is the ratio of correctly predicted negative samples to the total number of negative samples. PPV (positive predictive value) is the ratio of predicted positive samples to the total number of predicted positive samples. NPV (negative predictive value) is the ratio of predicted negative samples to the total number of predicted negative samples. These metrics reflect the model's predictive ability from different perspectives. Furthermore, the resulting metrics are compared with pre-set performance requirements. If all indicators meet the requirements, it means that the model performs well on the test set and has good generalization ability. The initial prediction model can be used as the target prediction model. If some indicators do not meet the requirements, it is necessary to readjust the model parameters or further process the data, and then repeat the training and evaluation process until the model performance meets the requirements.

[0108] In one feasible embodiment, the target prediction model can be constructed as a nomogram. A nomogram can clearly display the complex interrelationships between the various variables in the fusion model in an intuitive and visual manner. After successfully constructing the target prediction model, to comprehensively evaluate its performance and stability, the following method can be used: Calibration curves can be used to evaluate the accuracy of the model on both internal and external validation sets. The core function of the calibration curve is to reflect the degree of consistency between the model's predicted probability and the actual probability of occurrence. Specifically, a curve is plotted showing the correspondence between the model's predicted probability and the actual probability of occurrence to accurately assess the model's calibration. Ideally, the model's predicted probability should closely match the actual probability of occurrence, with the calibration curve approaching the diagonal. This indicates that the predicted probability given by the model truly reflects the likelihood of the actual event occurring, thus providing a reliable probabilistic basis for clinical decision-making. Decision curve analysis can be performed to evaluate the significant advantages of the net combined model in distinguishing recurrence risk categories and quantifying different probability thresholds. The key to decision curve analysis is to calculate the net benefit at different decision thresholds to further explore the model's practical application value in clinical decision-making scenarios. Specifically, under different threshold probability settings, the model's prediction results are compared with the net benefits of two extreme cases: no intervention measures and all intervention measures. This allows for a clear judgment of the specific circumstances under which the model can provide the greatest value for clinical decision-making, providing a strong reference for clinicians to rationally use the model under different risk judgments and decision thresholds. A stratified analysis was conducted based on multiple factors such as gender, age, and scanning equipment to verify the stability of the model in different category groups. Since patient groups of different genders and ages may have significant differences in their physiological characteristics and disease manifestations, and the data obtained by different scanning devices may also differ in quality, resolution, etc., these differences may affect the performance of the model. Through stratified analysis, we can fully understand the performance of the model under different populations and equipment conditions, ensure that the model has wide applicability and high reliability, thereby providing accurate and effective prediction services for different subgroups of patients and improving the application effect of the model in clinical practice.

[0109] In a feasible embodiment, after obtaining the target prediction model, the postoperative data of lung adenocarcinoma to be predicted can be input into the target prediction model to obtain a prediction result. Specifically, the patient's postoperative clinical data and CT images of lung adenocarcinoma can be input into the target prediction model. After receiving the input data, the target prediction model can analyze and calculate the data according to its internal algorithms and parameters. For the nomogram model, it can obtain a prediction result through a series of calculations and chart queries based on the pre-set relationship between each variable. The prediction result may be a specific numerical value (such as the probability of recurrence) or a classification result (such as high recurrence risk, low recurrence risk).

[0110] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the progression of lung adenocarcinoma after surgery, characterized in that: The method comprises: Acquire a clinical dataset and CT image data after surgery for lung adenocarcinoma, wherein the clinical dataset includes multiple clinical parameters and postoperative condition labels; Constructing a clinical data model according to the clinical data set, and obtaining an independent predictor based on the clinical data model, wherein the independent predictor is a parameter among the plurality of clinical parameters that has a predictive effect on a dependent variable of the clinical data model; constructing a radiomics feature model according to the CT image data, and obtaining feature weights based on the radiomics feature model; Constructing an image feature extraction model based on the CT image data and the pre-trained model, and obtaining a learning feature score based on the image feature extraction model; Constructing a target prediction model based on the independent predictor, the feature weights, and the learning feature scores; The postoperative data of lung adenocarcinoma to be predicted is input into the target prediction model to obtain a prediction result.

2. The method for predicting the progression of lung adenocarcinoma after surgery according to claim 1, wherein: The step of constructing a clinical data model based on the clinical data set and obtaining independent predictive factors based on the clinical data model includes: Modeling is performed based on each of the clinical parameters and the postoperative condition labels in the clinical data set to obtain a logistic regression model for each of the clinical parameters; Determining the relationship between the postoperative condition label and each of the clinical parameters according to a logistic regression model of each of the clinical parameters; Screening the clinical parameters whose relationship values ​​are less than a preset threshold from the clinical data set to obtain a set of predictive factors; Independent predictors are identified from the set of predictors.

3. The method for predicting the progression of lung adenocarcinoma after surgery according to claim 2, wherein: The identifying an independent predictor from the set of predictors comprises: constructing a first clinical data model according to the postoperative condition label; adding the predictor set to the first clinical data model, and calculating a first significance relationship value between the predictor set and the postoperative condition label; Filter out the prediction factor set corresponding to the first meaning relationship value being less than the preset threshold to obtain a first initial independent prediction factor; Modeling based on the set of predictors and the postoperative condition labels to obtain a second clinical data model; Calculating a second significance relationship value corresponding to each of the predictor sets in the second clinical data model, and filtering out the predictor sets whose second significance relationship values ​​are less than the preset threshold, until the second significance relationship values ​​corresponding to each of the predictor sets in the second clinical data model are all less than the preset threshold, thereby obtaining a second initial independent predictor; An independent predictor is confirmed based on the first initial independent predictor and the second initial independent predictor.

4. The method for predicting the progression of lung adenocarcinoma after surgery according to claim 1, wherein: The step of constructing a radiomics feature model based on the CT image data includes: Marking the lesion area on each CT image in the CT image data to obtain a key area image corresponding to each CT image; Extracting features from the key area image to obtain an image feature dataset; Performing feature screening processing on the image feature dataset to obtain a modeling feature dataset; Based on the modeling feature dataset, an imaging omics feature model is constructed.

5. The method for predicting the postoperative progression of lung adenocarcinoma according to claim 4, characterized in that: Obtaining feature weights based on the radiomics feature model includes: Extracting the regression coefficient of the modeling feature dataset from the radiomics feature model; The feature weight of each feature is obtained by performing calculation based on the absolute value of the regression coefficient.

6. The method for predicting the postoperative progression of lung adenocarcinoma according to claim 4, characterized in that: The performing feature screening processing on the image feature dataset to obtain a modeling feature dataset includes: Screening the image feature data set to obtain a preliminary feature data set; Performing one-way ANOVA on the preliminary feature data set to obtain an intermediate processing feature data set; A regression analysis is performed on the intermediate processed feature data set, and a modeling feature data set is obtained by screening according to the analysis results.

7. The method for predicting the progression of lung adenocarcinoma after surgery according to claim 1, characterized in that: The image feature extraction model is constructed based on the CT image data and the pre-trained model, including: resampling, lesion area marking, and stretching all CT images in the CT image data to obtain target processed images corresponding to each CT image; Adjusting parameters of the pre-trained model based on the CT image data to obtain a target feature extraction model; Using the target feature extraction model, performing feature extraction and feature screening processing on the target processed image corresponding to each of the CT images to obtain target image features; An image feature extraction model is constructed based on the target image features.

8. The method for predicting the postoperative progression of lung adenocarcinoma according to claim 7, characterized in that: The target feature extraction model is used to perform feature extraction and feature screening on the target processed images corresponding to each of the CT images to obtain target image features, including: merging target processed images corresponding to the CT images to obtain a plurality of target image blocks, each of the target image blocks including a first processed image, a second processed image, and a third processed image, wherein the first processed image corresponds to a first color channel, the second processed image corresponds to a second color channel, and the third processed image corresponds to a third color channel; Using the target feature extraction model, extracting features from the target image block to obtain preliminary image feature data; Dividing the preliminary image feature data into channels according to the first color channel, the second color channel, and the third color channel to obtain first channel feature data, second channel feature data, and third channel feature data; Extracting a first target feature from the first channel feature data, extracting a second target feature from the second channel feature data, and extracting a third target feature from the third channel feature data to obtain target channel feature data; Performing feature screening processing on each feature in the target channel feature data to obtain intermediate image feature data; The intermediate image feature data is subjected to feature screening processing by using ten-fold cross validation to obtain target image features.

9. The method for predicting the postoperative progression of lung adenocarcinoma according to claim 8, characterized in that: The obtaining of a learning feature score based on the image feature extraction model includes: Get the regression coefficient set of the image feature extraction model; Determining a regression coefficient of each feature in the target image feature according to the regression coefficient set; A learning feature score is calculated based on the regression coefficient of each feature.

10. The method for predicting the progression of lung adenocarcinoma after surgery according to claim 1, characterized in that: The constructing a target prediction model according to the independent predictor, the feature weight and the learning feature score includes: The independent predictors and learning feature scores are standardized, and the initial prediction model is constructed based on the processed independent predictors, learning feature scores, and feature weights; Get the prediction training set and prediction test set; The initial prediction model is trained according to the prediction training set, and the weight parameters of the initial prediction model are adjusted according to a preset error index, and the initial prediction model is obtained when the maximum number of iterations is reached or the error converges to a preset threshold; The initial prediction model is evaluated using the prediction test set, and when the performance of the initial prediction model in the prediction test set meets the requirements, the initial prediction model is used as the target prediction model.

Images