Multi-modal score construction method for identifying benign and malignant pulmonary nodules and application of multi-modal score construction method

Abstract

The invention discloses a multi-modal score construction method for identifying benign and malignant pulmonary nodules and application of the multi-modal score construction method. The multi-modal score construction method comprises the steps that 1, data are collected, and data sets serve as a training set and a verification set respectively; step 2, preprocessing the collected low-dose spiral CT image; 3, aiming at the preprocessed low-dose spiral CT image, extracting marker features of the pulmonary nodule under imaging; the pulmonary nodule deep learning malignant probability is acquired, and the pulmonary lobe radiomics detection result of the participant is acquired; 4, based on the extracted pulmonary nodule features, constructing a single-mode nodule score, and verifying and evaluating a score identification effect; and step 5, constructing a pulmonary nodule benign and malignant identification multi-mode comprehensive score. According to the method, after the patient completes low-dose spiral CT or provides biological information, the individualized risk score is automatically generated, and the relative risk level in the screening crowd is visually displayed in a risk percentage mode to assist doctors in decision making and grading intervention.

Description

TECHNICAL FIELD

[0001] The application relates to a multi-modal scoring method for differentiating between benign and malignant pulmonary nodules and application thereof, and belongs to the technical field of biomedical information engineering. BACKGROUND

[0002] Low-dose computed tomography (LDCT) has been recognized as the only effective lung cancer screening method. Lung cancer screening exists in the form of pulmonary nodules on CT images. With the development of imaging technology, the detection rate of pulmonary nodules has also increased year by year.

[0003] The existing nodule management guidelines mainly classify nodules according to the density and size of the nodules to further guide the clinical judgment of the benign and malignant pulmonary nodules. The existing several international mainstream lung nodule benign and malignant differentiation models such as Mayo model, Brock model and the like integrate clinical risk factors and nodule imaging features to evaluate the malignancy degree of the nodules. However, a plurality of studies have found that the construction of the nodule feature prediction model of the benign and malignant nodules still has the problems of limited sample size, no effect evaluation in a large-scale screening cohort and the like. Meanwhile, the above-mentioned models have a single variable dimension, and often ignore the deep and systematic integration of a plurality of potential risk factors such as high-throughput imaging features and individual genetic risks.

[0004] With the wide application of artificial intelligence (AI) and deep learning (DL) in the field of medical imaging, the automatic feature extraction based on AI provides a more efficient and objective solution for the risk assessment of pulmonary nodules. However, the current lung nodule benign and malignant evaluation in lung cancer screening has the following key technical problems:

[0005] Technical problem one: the single variable dimension is difficult to reflect the complexity of the lesion;

[0006] Most of the existing guidelines and differentiation models are based on traditional variables (such as age, gender, smoking history, nodule size and the like), and cannot fully integrate the high-dimensional data available in modern medicine, especially the potential effective indicators such as imaging features, deep learning features and genetic information, so that the discrimination and stability of the risk assessment are limited.

[0007] Technical problem two: the feature extraction depends on manual work, has strong subjectivity and low efficiency;

[0008] Most current models rely on clinicians for manual reading and feature extraction, which has the problems of inconsistent evaluation standards, large subjective errors, and difficulty in large-scale screening applications. Manual reading and feature extraction not only have low efficiency, but also are difficult to ensure the consistency and repeatability of data quality.

[0009] Technical problem three: the existing model has limited effect on the Chinese population;

[0010] Most existing models are derived from European and American populations, and do not fully consider the unique nodular distribution characteristics, epidemiology and genetic background differences of the Chinese population. When directly applied to the Chinese population, the accuracy and applicability are limited, which limits its promotion and application value. Currently, there is still a lack of lung nodule benign and malignant scoring tools based on large sample Chinese population.

[0011] Technical problem four: lack of model stability and generalization ability, lack of systematic verification;

[0012] Most studies construct models based on a single data source or case sample, lack of verification based on large sample, multi-center population, and cannot fully reflect the stability and promotion potential of the model in real screening scenarios.

[0013] Therefore, it is urgent to construct a lung nodule benign and malignant discrimination method that integrates multi-source information and is suitable for the Chinese population to improve screening efficiency and support the ability to assist clinical decision-making. SUMMARY

[0014] The purpose of the present application is to provide a multi-modal scoring construction method for lung nodule benign and malignant discrimination and its application. The method integrates multi-modal information, including imaging features of lung nodules, deep learning malignancy probability, population epidemiology characteristics, single nucleotide polymorphism (SNPs) detection results, and lung lobe radiomics features, to achieve high-precision comprehensive scoring of lung nodules. The relative risk level in the screening population is intuitively displayed through the risk percentage method, which assists doctors in decision-making and graded intervention.

[0015] To achieve the above purpose, the technical solution adopted by the present application is: a multi-modal scoring construction method for lung nodule benign and malignant discrimination, which comprises the following steps:

[0016] Step 1, data collection: collect case information of benign and malignant lung nodules with clear pathological diagnosis, and collect population epidemiology information of participants in a prospective lung cancer screening cohort, and collect SNPs detection results of participants in the validation data set, and collect chest low-dose spiral CT images of participants in the two data sets;

[0017] Step 2, data preprocessing: pre-process the collected low-dose spiral CT images;

[0018] Step 3, feature extraction: for the pre-processed low-dose spiral CT image, the marker feature of the lung nodule in the imaging is extracted; and the lung nodule deep learning malignancy probability is obtained, and the lung lobe imageomic detection result of the participant is obtained;

[0019] Step 4, constructing and verifying single-mode nodule score: based on the extracted lung nodule features, a single-mode nodule score is constructed, and the identification effect of the score is evaluated in the verification data set;

[0020] Step 5, constructing a multi-modal comprehensive score: based on the single-mode nodule score constructed in step 4, and the participant's population epidemiological characteristics, SNPs detection results collected in step 1, and the lung nodule deep learning malignancy probability, lung lobe imageomic detection result collected in step 3, the Cox proportional hazards model and the ridge regression model are used for feature screening and weighting to construct a multi-modal comprehensive score for identifying lung nodule benign and malignant.

[0021] Further, in step 1, the process of data collection includes:

[0022] The training data comes from lung nodule patients with clear pathological diagnosis, including benign and malignant nodules, and all cases complete LDCT scanning within 1 month before operation, the inclusion criteria include: no history of lung surgery, exclude other tumor metastasis and preoperative intervention; The benign and malignant participants are matched according to gender and age in a 1:1 ratio, and a certain number of cases are constructed as a training set;

[0023] The verification data set includes lung cancer tissue screening cohort A and lung cancer opportunistic screening cohort B, the inclusion criteria of cohort A: smokers need to meet the following conditions: smoking pack years ≥20, including those who have smoked ≥20 pack years and have quit smoking for less than 15 years; Non-smokers are considered high risk if they meet any of the following conditions:

[0024] (1) Non-smoking women who have been passively smoking for ≥20 years living together or working in the same room with smokers who meet the above conditions,

[0025] (2) First-degree relatives with lung cancer family history,

[0026] (3) Exposure to carcinogens related to occupation for ≥1 year,

[0027] (4) History of chronic obstructive pulmonary disease;

[0028] The inclusion criteria of cohort B is: participants aged 18 years and older who are suitable for low-dose spiral CT; Collect the participant's population epidemiological information, and collect the participant's SNPs detection results.

[0029] Further, in step 2, the process of data preprocessing includes: resampling all low-dose spiral CT images to isotropic 1x1x1mm 3Voxel spacing; image intensity normalization in standard lung window; batch effect correction using ComBat method; automatic segmentation of all lung nodules and five lung lobes.

[0030] Further, in step 3, the feature extraction includes:

[0031] (1) Nodule image features: 33 features are extracted, including 18 continuous variables (long axis, short axis, average diameter, 3D axis length, volume, mass, surface area, maximum layer area, nodule height, sphericity, compactness, CT maximum value, CT minimum value, CT average value, entropy, skewness, kurtosis, standard deviation) and 15 categorical variables (location (upper / middle / lower lobe), density (solid / glassy / partial solid), burr sign, cavity, spinous process, pleural traction, edge, lobulation, air bronchogram, vascular inclusion, vascular bundle);

[0032] (2) Deep learning malignancy probability: calculate the AI malignancy of the nodule risk probability;

[0033] (3) Lung lobe radiomics features: based on the lung lobe segmentation results completed in step 2, the imageomics features of the lung lobe in the region of interest are extracted using the PyRadiomics toolkit.

[0034] Further, in step 4, the process of constructing and verifying the single-mode nodule score includes: based on the nodule features extracted in step 3, using the R language glmnet package to perform LASSO fitting on the nodule features; using ten-fold cross-validation to select the optimal regularization parameter λ on the training set; the final feature selection is based on lambda.min, that is, the λ value that makes the model cross-validation error minimum, to ensure that the most discriminative variables are retained while controlling the model complexity.

[0035] Further, according to the selected features and their coefficients by LASSO fitting, a single-mode nodule scoring system is established, and the performance is evaluated on an independent validation set; the specific expression of the single-mode nodule score is as follows:

[0036] Node Score=

[0037] -18.40994-0.48446×location_l+0.11543×location_c+0.33998×density_ggo

[0038] -1.32912 x density_solid + 0.01128 x major_diameter - 0.00866 x minor_diameter - 0.0000224 x volume - 0.000526 x hu_max - 0.00274 x hu_min + 0.00114 x hu_mean + 0.00096 x max_area_mm - 0.03487 x maximum_3d_diameter + 4.82030 x sphericity + 1.06405 x entropy + 0.22880 x skewness - 0.12753 x kurtosis - 0.00935 x standard_deviation + 0.09891 x height_diameter - 1.85447 x lung_spiculation - 0.41264 x cavitation + 5.35330 x ratchet_protrusion + 1.79332 x tractive_pleural + 0.40887 x margin + 0.75031 x lung_lobulation + 0.65262 x air_bronchogram + 0.48514 x vascular_embedding - 1.08487 x vascular_notch

[0039]

[0040] wherein location_l and location_c represent the location of the nodule in the lung lobe lower lobe or middle lobe; density_ggo and density_solid represent the nodule as pure ground glass opacity or solid density; major_diameter represents the longest diameter of the nodule measured in the transverse plane, in millimeter mm; minor_diameter represents the shortest diameter perpendicular to the longest diameter, in millimeter mm; volume is the total volume of the nodule in three-dimensional space, in mm 3 ; hu_max, hu_min and hu_mean represent the maximum, minimum and average CT values in the nodule region, respectively, in Hounsfield Unit; max_area_mm represents the maximum area of the nodule in the two-dimensional cross section, in mm 2 ​; maximum_3d_diameter represents the longest diameter of the nodule in three-dimensional space, in millimeters (mm); sphericity represents the degree of sphericity of the nodule, and the closer the value is to 1, the closer the nodule is to a sphere; entropy is a complexity index of gray scale distribution, and the greater the value, the more complex the texture; skewness and kurtosis measure the skewness and kurtosis of the gray scale distribution, respectively; standard_deviation represents the standard deviation of the CT value of the nodule; height_diameter represents the vertical height of the nodule in the longitudinal direction of the lung nodule, in millimeters (mm); lung_spiculation represents the spiculation sign; cavitation represents the cavitation sign; ratchet_protrusion represents the ratchet protrusion; tractivepleural represents the traction pleural sign; margin represents the blurred edge; lung_lobulation represents the lobulation sign; air_bronchogram represents the air bronchogram; vascular_embedding represents the vascular embedding sign; and vascularnotch represents the vascular notch sign.

[0041] Further, in step 5, the process of constructing the multi-modal nodule score includes: on the basis of the single-modal score, further integrating the nodule malignancy probability output by deep learning, population epidemiological characteristics, SNPs, and lung lobe image features; screening the single-factor correlation P < 0.2 of the predictive factors by the Cox proportional risk model, and using the coefficients output by the ridge regression to perform weighted modeling, to establish a multi-modal comprehensive discriminant score for the benign and malignant lung nodules.

[0042] Still further, independent multi-modal comprehensive scores are constructed for smokers and non-smokers, respectively, and the specific expressions of the comprehensive scores are as follows:

[0043] Multi-dim score smoker =

[0044] 0.35 x Node score + 3.64 x AI malignance + 7.41 x 10 -3 PRS + 2.29 x 10 -2 Radiomic score;

[0045] Multi-dim score nonsmoker =

[0046] 0.48 x Node score + 3.84 x AI malignance + 0.57 x PRS + 1.56 x 10 -2Radiomicscore;

[0047] wherein Node score is the single modality score built in step 4, AI malignance is the probability of lung nodule malignancy predicted by deep learning model, LCRS score is the lung cancer epidemiology risk score of the participant, PRS is the lung cancer polygenic risk score, and Radiomic score is the imageomics score at the lung lobe level.

[0048] Further, the calculation method of LCRS score is to set the regression coefficient and score value of each variable according to the preset reference group, and add the score results of all variables to obtain the total LCRS epidemiology risk score.

[0049] PRS is the lung cancer polygenic risk score, which comes from the genotype data of each SNP site extracted in step 3. For each SNP site, a numerical value (i.e. site score) is assigned according to the chip detection result: wild homozygous type is recorded as 0, heterozygous type is recorded as 1, and variant homozygous type is recorded as 2; the genotype numerical value of each SNP is substituted into the PRS formula, and finally the PRS value is equal to the sum of the product of each site genotype numerical value and the corresponding weight coefficient. The specific calculation formula of PRS is: PRS = (0.1016 x score of rs17038564) + (0.1174 x score of rs2293607) + (0.1900 x score of rs11375254) + (0.2653 x score of rs13167280) + (0.1447 x score of rs401681) + (0.1826 x score of rs2517873) + (0.0614 x score of rs3817963) + (0.1399 x score of rs1853837) + (0.0843 x score of rs5879422) + (0.1555 x score of rs4236709) + (0.1004 x score of rs10429489) + (0.0949 x score of rs35201538) + (0.0896 x score of rs4573350) + (0.1338 x score of rs12265047) + (0.1184 x score of rs55768116) + (0.0648 x score of rs11610143) + (0.1002 x score of rs1200399) + (0.0771 x score of rs77468143) + (0.1144 x score of rs200595745).

[0050] Radiomic score is the imageomics score at the lung lobe level, i.e. left upper lobe UL, left lower lobe LL, right upper lobe UR, right middle lobe CR and right lower lobe LR. The calculation method of lung lobe imageomics score is the weighted sum of the product of each imageomics variable and weight.

[0051] Further, after the above-mentioned multi-modal score is constructed, the discrimination ability of the model is evaluated with AUC as the main index; meanwhile, sensitivity, specificity, accuracy, positive predictive value, negative predictive value and other indexes are calculated to comprehensively measure the performance of the model compared with other mainstream models; and the prediction gain of the multi-modal score compared with the single-modal score is further evaluated through the net reclassification index (NRI).

[0052] An application of the multi-modal score construction method in the lung nodule risk discrimination, the score constructed is used for assisting clinical decision-making in the actual lung cancer screening population, and optimizing the lung cancer early screening process.

[0053] Compared with the prior art, the present application has the following advantages:

[0054] The score of the present application can fully integrate nodule image features, deep learning recognition results, population epidemiological characteristics, genetic information and lung lobe imageomics features, and embodies the advantages of the multi-modal integration strategy of the present application. The score system shows high discrimination ability in the verification queue, and is suitable for risk stratification and individualized diagnosis support in the real lung cancer screening scene. Compared with the traditional lung nodule benign and malignant discrimination model, the present application has the following technical advantages:

[0055] (1) Batch automatic feature extraction based on an artificial intelligence platform: the present application realizes batch automatic extraction of lung nodule image features based on an artificial intelligence platform, overcomes the subjectivity and low efficiency problems caused by traditional manual film reading, and guarantees the objectivity and consistency of feature extraction.

[0056] (2) Fusion of multi-dimensional prediction factors to improve discrimination accuracy: the present application fuses lung nodule image features, deep learning malignancy probability, population epidemiological information, genetic risk score and lung lobe level imageomics features and other multi-dimensional prediction factors to construct a comprehensive evaluation model, which significantly improves the accuracy of lung nodule benign and malignant risk discrimination.

[0057] (3) Designed for Chinese population to improve applicability and application value: the training set and verification set used in the present application are based on Chinese population, which overcomes the limitation of the existing Mayo model, Brock model and other models based on European and American population in recognizing the efficiency of Chinese population, and improves the population adaptability and practical application value of the model.

[0058] (4) Rigorous design to ensure model stability and promotion potential: The present application adopts a three-stage rigorous design, covering case-control data set construction and two independent lung cancer screening cohorts, to ensure the stability and reliability of the model in different populations and scenarios. This design not only verifies the scientificity of the model, but also lays a solid foundation for its promotion and application in actual screening sites.

[0059] In summary, based on the case-control and large-scale lung cancer screening population design, the present application focuses on the identification of benign and malignant lung nodules, through standardized image preprocessing process and automatic extraction of nodule features, and systematically integrates nodule image features to construct a single modality score for identifying benign and malignant nodules. Further, a multi-dimensional comprehensive score for identifying benign and malignant lung nodules suitable for Chinese population is constructed by using machine learning algorithm to integrate multi-source data. The score has been verified in a prospective screening cohort and shows superior discrimination ability, which can provide scientific basis for accurate identification and individualized management of high-risk population of lung cancer. The research results not only break through the limitations of traditional image size and density evaluation, but also provide key technical support for promoting the intelligentization, precision and standardization of lung cancer screening in China, and have important clinical application value and public health significance. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is a flowchart of the lung nodule benign and malignant classification identification score method of an embodiment of the present application.

[0061] Figure 2 is a parameter selection process diagram for constructing a nodule score using LASSO, and a score distribution of a single modality nodule score in case-control data set, verification data set 1, and verification data set 2 participants.

[0062] Figure 3 is a ROC, sensitivity, and specificity comparison diagram of a single modality score and Mayo, Brock, VA, PKUPH, and different models for distinguishing high-risk population with 6mm as the cutoff in 20565 verification cohort 1 and 5703 verification cohort 2 data sets.

[0063] Figure 4 is a ROC curve comparison diagram of multi-modality score, single modality score, and Mayo, Brock, VA, PKUPH, and different models for distinguishing high-risk population with 6mm as the cutoff in 10576 smoking and 8180 non-smoking training sets. DETAILED DESCRIPTION

[0064] The present application will be described in detail below with reference to the accompanying drawings.

[0065] A multi-modality score construction method for identifying benign and malignant lung nodules includes the following steps:

[0066] Step 1, data collection: collect the information of patients with benign and malignant lung nodules with clear pathological diagnosis, and the epidemiological information of participants in the prospective lung cancer screening cohort, and the two data sets are respectively as training set and validation set, collect the SNPs detection results of the participants in the validation data set, and collect the chest low-dose spiral CT (LDCT) images of the participants in the two data sets.

[0067] The process of data collection includes: the training data comes from lung nodule patients with clear pathological diagnosis, including benign and malignant nodules, and all cases complete LDCT scan within 1 month before operation. The inclusion criteria include: no history of lung surgery, exclude other tumor metastasis and preoperative intervention (radiotherapy, puncture, etc.) in malignant cases. The benign and malignant participants are matched according to gender and age 1:1, and 935 pairs of cases are constructed as training set; The validation data set includes lung cancer tissue screening cohort 1 and lung cancer opportunistic screening cohort 2. The inclusion criteria of cohort 1 refers to the national major public health service project, which requires that smokers meet the following conditions: smoking pack years ≥ 20, including those who have smoked ≥ 20 pack years and have quit smoking for less than 15 years; Non-smokers are judged as high risk if they meet any of the following conditions: (1) non-smoking women who live together or work in the same room with smokers who meet the above conditions have been passively smoking for more than 20 years, (2) family history of lung cancer in first-degree relatives (including parents, children and siblings), (3) exposure to carcinogens related to occupation (including asbestos, radon, beryllium, chromium, cadmium, silicon, coal smoke and coal soot) for more than 1 year, (4) history of chronic obstructive pulmonary disease (COPD). The inclusion criteria of cohort 2 is: participants over 18 years old who are suitable for low-dose spiral CT (LDCT). Collect the participants' epidemiological information, including age, education level, height, body mass index BMI, whether to exercise regularly, whether to cough frequently, whether to have emphysema or bronchitis, whether to have a personal history of cancer, whether to have a history of cancer in two or more first-degree relatives, daily smoking amount, smoking years, whether to inhale smoke deeply into the lungs when smoking, and the length of time since quitting smoking. Collect the SNPs detection results of the participants, mainly through gene chip to obtain the genetic information of lung cancer related SNPs sites, and the specific SNPs site information is shown in Table 1.

[0068] Table 1. Construction of PRS related SNPs site

[0069]

[0070]

[0071] Step 2, data preprocessing: based on LDCT image for pretreatment.

[0072] The process of data preprocessing includes: resample all LDCT images to isotropic 1x1x1mm 3Voxel spacing; image intensity normalization in standard lung window; batch effect correction using ComBat method. In addition, all lung nodules and five lobes of the lung were automatically segmented using a deep learning artificial intelligence system developed by Beijing Deep Radiomics Technology Co., Ltd. All LDCT images were subjected to the above standardization preprocessing procedures before extracting nodule features and lobar radiomic features to minimize inter-center variability due to differences in scanner models and imaging protocols.

[0073] Step 3, feature extraction: for the preprocessed LDCT images, extract the marker features of lung nodules in imaging; obtain the deep learning malignancy probability of lung nodules; obtain the lobar imaging results of participants.

[0074] The process of feature extraction includes: ①Nodule imaging features: 34 features are automatically extracted based on the deep learning artificial intelligence system developed by Beijing Deep Radiomics Technology Co., Ltd., including size (maximum diameter, minimum diameter, average diameter), density (CT value, density type), location, three-dimensional structure (volume, surface area), etc. The specific nodule features are shown in Table 2; ②Deep learning malignancy probability: use the deep learning model trained by the above artificial intelligence system to calculate the malignancy risk probability AImalignance of the nodule; ③Lobar radiomic features: based on the lobar segmentation results completed in step 2, use the PyRadiomics toolkit to extract the imaging features of the lobe within the region of interest (ROIs); including texture, shape, intensity and other features of the lobe.

[0075] Table 2. Lung nodule features and variable types

[0076]

[0077]

[0078] Step 4, construct and verify single-modality nodule score: based on the extracted lung nodule features, use the least absolute shrinkage and selection algorithm (LASSO) to construct a single-modality nodule score, and evaluate the scoring effect in the validation data set.

[0079] The process of constructing and validating the single modality nodule score includes: based on the nodule features extracted in step 3, using the R language glmnet package to perform LASSO fitting on the logistic regression model. The optimal regularization parameter λ is selected on the training set in the way of 10-fold cross-validation. The final feature selection is based on lambda.min, that is, the λ value that makes the model cross-validation error minimum, to ensure that the most discriminative variables are retained while controlling the model complexity. According to the features and their coefficients selected by LASSO regression, a single modality nodule score system is established, and the performance is evaluated on the independent validation set. The specific expression of the single modality nodule score is as follows: Node Score=-18.40994-0.48446×location_l+0.11543×location_c+0.33998×density_ggo-1.32912×density_solid+0.01128×major_diameter-0.00866×minor_diameter-0.0000224×volume-0.000526×hu max-0.00274×hu min+0.00114×hu mean+0.00096×max area mm-0.03487×maximum_3d_diameter+4.82030×sphericity+1.06405×entropy+0.22880×skewness-0.12753×kurtosis-0.00935×standard_deviation+0.09891×height_diameter-1.85447×lung_spiculation-0.41264×cavitation+5.35330×ratchet_protrusion+1.79332×tractive_pleural+0.40887×margin+0.75031×lung_lobulation+0.65262×air_bronchogram+0.48514

[0080] ×vascular_embedding-1.08487×vascular_notch;

[0081] wherein location_l and location_c represent the nodule is located in the lower lobe or middle lobe of the lung; density_ggo and density_solid represent the nodule is pure ground-glass density or solid density; major_diameter represents the longest diameter of the nodule measured in the cross-section, in millimeter (mm); minor_diameter represents the shortest diameter perpendicular to the longest diameter, in millimeter (mm); volume is the total volume of the nodule in three-dimensional space (unit mm 3 ); hu_max, hu_min and hu_mean represent the maximum, minimum and mean CT value (unit Hounsfield Unit) in the nodule region, respectively; max_area_mm represents the maximum area of the nodule in the two-dimensional cross-section (unit mm 2 ); maximum_3d_diameter represents the longest diameter of the nodule in three-dimensional space, in millimeter (mm); sphericity represents the degree of sphericity of the nodule, the closer to 1, the closer to spherical; entropy is the complexity index of gray distribution, the greater the value, the more complex the texture; skewness and kurtosis measure the skewness and kurtosis of the gray distribution, respectively, which are helpful to capture the asymmetric and centralized trend characteristics of the nodule texture distribution; standard_deviation represents the standard deviation of the nodule CT value, reflecting the uneven degree of its density; height_diameter represents the vertical height of the nodule in the longitudinal direction of the lung nodule, in millimeter (mm); in addition, lung_spiculation (spiculation sign), cavitation (cavitation sign), ratchet_protrusion (ratchet protrusion), tractivepleural (traction of the pleura), margin (blurred edge), lung_lobulation (lobulation), air_bronchogram (air bronchogram), vascular_embedding (vascular embedding) and vascularnotch (vascular notch) are the imaging signs commonly used by radiologists to judge the benign and malignant of lung nodules, and the above variables jointly constitute the input features of the single-modality nodule score. The evaluation indexes of the single-modality score include area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value, negative predictive value, etc., to measure the benign and malignant discrimination ability of the score.

[0082] Step 5, constructing multi-modal comprehensive score: based on the single-modal nodule score constructed in step 4, and the participant's population epidemiological characteristics, SNP detection results collected in step 1, and the lung nodule deep learning malignancy probability collected in step 3, the Cox proportional risk model and the ridge regression model are used for feature screening and weighting to construct a multi-modal comprehensive score for identifying lung nodule benignity and malignancy.

[0083] The process of constructing a multi-modal comprehensive score includes: on the basis of a single-modal score, further integrating the other four types of features (nodule malignancy probability output by deep learning, population epidemiological characteristics, SNPs, and lung lobe imageomics features). Through the Cox proportional risk model, the single-factor correlation P<0.2 is screened out as a predictive factor, and the ridge regression (Ridge Regression) output coefficient is used for weighted modeling to establish a multi-modal comprehensive score for identifying lung nodule benignity and malignancy. In order to consider the differences in lung cancer pathogenesis between smokers and non-smokers, the present application constructs independent multi-modal comprehensive scores for smokers and non-smokers, respectively, and the specific expressions of the comprehensive scores are as follows:

[0084] Multi-dim score smoker =0.35×Node score+3.64×AI malignance+7.41×10 -3 ×LCRS score+0.11×PRS+2.29×10 -2 ×Radiomic score;

[0085] Multi-dim score nonsmoker =0.48×Node score+3.84×AI malignance+0.57×PRS+1.56×10 -2 ×Radiomic score,

[0086] Wherein Node score is the single-modal score constructed in step 4; AI malignance is the lung nodule malignancy probability predicted by the deep learning model, LCRS score is the lung cancer epidemiological risk score of the participant, PRS is the lung cancer multi-gene risk score, and Radiomic score is the imageomics score at the lung lobe level.

[0087] Specifically, the calculation of the AI malignance score is based on a deep learning model trained by Beijing Shengli Bo United Technology Co., Ltd., which is used to predict the probability of lung nodules being malignant. The model adopts a convolutional neural network architecture, combining the structural advantages of residual network (ResNet) and efficient network (EfficientNet), and can automatically extract multi-scale spatial features from LDCT images to realize the classification and prediction of nodules. The input of the model is the standardized three-dimensional image block of the nodule, and the output is the malignant probability score of the nodule (range 0-1), the higher the value, the higher the risk of malignancy.

[0088] Specifically, the calculation method of LCRS score is to set the regression coefficient and score value of each variable according to the preset reference group, and add the score results of all variables to obtain the total LCRS epidemiological risk score. The specific risk score assignment is as shown in Table 3:

[0089] Table 3. Risk score assignment table

[0090]

[0091]

[0092]

[0093] Specifically, the PRS is a lung cancer polygenic risk score, which is derived from the genotype data of each SNP site in step 3 feature extraction. For each SNP site, a numerical value (i.e., site score) is assigned according to the chip detection result: wild homozygous type is recorded as 0, heterozygous type is recorded as 1, and variant homozygous type is recorded as 2; the genotype numerical value of each SNP is substituted into the PRS formula, and finally the PRS value is equal to the sum of the product of each site genotype numerical value and the corresponding weight coefficient. The specific calculation formula of PRS is: PRS = (0.1016 x score of rs17038564) + (0.1174 x score of rs2293607) + (0.1900 x score of rs11375254) + (0.2653 x score of rs13167280) + (0.1447 x score of rs401681) + (0.1826 x score of rs2517873) + (0.0614 x score of rs3817963) + (0.1399 x score of rs1853837) + (0.0843 x score of rs5879422) + (0.1555 x score of rs4236709) + (0.1004 x score of rs10429489) + (0.0949 x score of rs35201538) + (0.0896 x score of rs4573350) + (0.1338 x score of rs12265047) + (0.1184 x score of rs55768116) + (0.0648 x score of rs11610143) + (0.1002 x score of rs1200399) + (0.0771 x score of rs77468143) + (0.1144 x score of rs200595745), and the size of the PRS numerical value represents the susceptibility of the individual at the genetic level to lung cancer.

[0094] Specifically, the Radiomic score is an imageomics score at the lung lobe level, i.e., left upper lobe (UL), left lower lobe (LL), right upper lobe (UR), right middle lobe (CR), and right lower lobe (LR). The calculation method of the lung lobe imageomics score is the weighted sum of the product of each imageomics variable and the weight. The specific imageomics score includes variables and weights as shown in Table 4:

[0095] Table 4. Variables and weights involved in imageomics score

[0096]

[0097]

[0098]

[0099] After constructing the above multi-modal score, the AUC is taken as the main index to evaluate the discriminant ability of the model. At the same time, the sensitivity, specificity, accuracy, positive predictive value, negative predictive value and other indicators are calculated to comprehensively measure the performance of the model compared with other mainstream models. The prediction gain of the multi-modal score compared with the single-modal score is further evaluated by the net reclassification index (NRI).

[0100] The above method is further described below based on specific embodiments:

[0101] Example 1: Construction, verification and comparison with mainstream models of single-modal score

[0102] First, LASSO regression method is used for variable screening. The optimal regularization parameter λ value is determined by 10-fold cross-validation, and the λ (lambda.min) that makes the average error minimum is selected as the final modeling parameter. On this basis, the key variables with non-zero regression coefficients are screened out, and the corresponding regression coefficients are combined to calculate the single participant's nodule score (Node Score). Figure 1 The specific method is: multiply each nodule feature by its corresponding LASSO regression coefficient and sum up to get the final single-modal score. The final score formula is as follows:

[0103] Node Score=

[0104] -18.40994-0.48446×location_l+0.11543×location_c+0.33998×density_ggo

[0105] -1.32912×density_solid+0.01128×major_diameter-0.00866×minor_diameter-0.0000224×volume-0.000526×hu_max-0.00274×hu_min+0.00114×hu_mean+0.00096×max_area_mm-0.03487×maximum_3d_diameter+4.82030×sphericity+1.06405×entropy+0.22880×ske wness-0.12753×kurtosis-0.00935×standard_deviation+0.09891×height_diameter-1.85447×lung_spiculation-0.41264×cavitation+5.35330×ratchet_protrusion+1.79332×tractivepleur al+0.40887×margin+0.75031×lung_lobulation+0.65262×air_bronchogram+0.48514

[0106] ×vascular_embedding-1.08487×vascularnotch;

[0107] wherein location_l and location_c represent the nodule is located in the lower lobe or middle lobe of the lung; density_ggo and density_solid represent the nodule is pure ground-glass density or solid density; major_diameter represents the longest diameter of the nodule measured in the cross-sectional plane, in millimeters (mm); minor_diameter represents the shortest diameter perpendicular to the longest diameter, in millimeters (mm); volume is the total volume of the nodule in three-dimensional space (unit mm3); hu_max, hu_min and hu_mean represent the maximum, minimum and average CT values (in Hounsfield Unit) in the nodule region, respectively; max_area_mm represents the maximum area of the nodule in the two-dimensional cross-sectional plane (unit mm2); maximum_3d_diameter represents the longest diameter of the nodule in three-dimensional space, in millimeters (mm); sphericity represents the degree of sphericity of the nodule, and the closer the value is to 1, the closer the nodule is to a sphere; entropy is a complexity index of gray scale distribution, and the greater the value, the more complex the texture; skewness and kurtosis measure the skewness and kurtosis of the gray scale distribution, respectively, which are helpful to capture the asymmetric and centralized trend characteristics of the nodule texture distribution; standard_deviation represents the standard deviation of the nodule CT value, reflecting the unevenness of its density; height_diameter represents the vertical height of the nodule in the longitudinal direction of the lung nodule, in millimeters (mm); in addition, lung_spiculation (spiculation sign), cavitation (cavitation sign), ratchet_protrusion (ratchet protrusion), tractivepleural (traction of the pleura), margin (blurred edge), lung_lobulation (lobulation), air_bronchogram (air bronchogram), vascular_embedding (vascular embedding), and vascularnotch (vascular notch) are the imaging signs commonly used by radiologists to judge the benign and malignant lung nodules. The above variables collectively constitute the input features of the single-modality nodule score, which are used to comprehensively characterize the image phenotype of the lung nodule.

[0108] According to the above single-modality nodule score calculation method, the calculation of the nodule score was performed on two validation data sets. With the new lung cancer as the outcome of the nodule malignant transformation during the follow-up of the cohort, the Cox proportional hazards model was used to evaluate the application effect of the nodule score, and the ROC was further drawn to evaluate the sensitivity and specificity of the prediction, and then to evaluate the ability of the score constructed by these nodule characteristics to distinguish benign and malignant lung nodules (Table 5, Figure 2Results showed that the AUC of the score constructed by 27 nodular features could reach 0.905 and 0.870 in two validation centers, and the sensitivity could reach 0.785 and 0.828, and the specificity was 0.867 and 0.762, respectively. Compared with Mayo, Brock, VA, PKUPH and different models with 6mm as the cutoff to distinguish high-risk population, the score achieved the best or better discrimination effect in both validation centers.

[0109] Table 5 Specific performance index data of single modality score and each mainstream model in the validation data set

[0110]

[0111]

[0112] Combined with the screening epidemiological data and lung cancer incidence data (Tables 6 and 7), it can be found that the single modality score constructed in this study has good risk stratification ability. The specific performance is as follows: in the population, the highest 20% score group was found to be concentrated in most lung cancer cases, in which: in the validation queue 1 data set, the highest 20% score group contained 74.89% of lung cancer cases, while the lowest 20% score group accounted for only 1.5%. The cumulative detection rate of lung cancer increased significantly in the high score group, showing a significant decreasing trend (P<0.001); in the validation queue 1 data set, 87.68% of lung cancer cases in the highest 20% score group were detected within one year, indicating that the high score population has a more significant short-term risk of nodular malignancy, which suggests its importance in the early screening stage of lung cancer. In the validation queue 2 data set, the highest 20% score group also concentrated 75% of lung cancer cases, while the last three score groups almost had no case detection, which also indicates that the score system has good discrimination ability in the general population.

[0113] Table 6 Detection of single modality score in validation queue 1 data set

[0114]

[0115] Table 7 Detection of single modality score in validation queue 2 data set

[0116]

[0117] Example 2: Construction and evaluation of multi-modality comprehensive discriminant score

[0118] On the basis of the foregoing single-modal nodule score, in order to further improve the discrimination ability of the lung nodule benign and malignant risk, the application integrates multi-dimensional key feature information, including a lung cancer risk score (LCRS score) based on epidemiological characteristics, an AI malignancy probability score (AI malignance) output by a deep learning model, a polygenic risk score (PRS), and a lobe-level imageomics score (Radiomic score), establishes a multi-modal comprehensive score, and constructs a stratified score according to the smoking status.

[0119] Firstly, based on the actual nodule outcome in the verification data set 1, the foregoing five types of features are taken as input variables, and Cox proportional risk model and ridge regression (Ridge Regression) are used to jointly model and optimize the weight of each type of feature. In the Cox proportional risk model analysis, firstly, single-factor analysis is performed on each candidate feature, and variables with potential prediction value (P value <0.2) are screened out, and then the screened variables are included in the ridge regression model for further analysis. The Cox proportional risk model can fully consider the time factor and model the risk of benign and malignant outcomes, and the ridge regression can effectively solve the problem of multicollinearity by introducing an L2 regularization term, and stabilize the regression coefficient estimate, thereby improving the discrimination efficiency of the comprehensive score.

[0120] The finally formed score is stratified according to the smoking status, and multi-modal comprehensive scores suitable for smokers and non-smokers are constructed respectively. The specific formula is as follows:

[0121] For smokers, the comprehensive score is Multi-dim Score Smoker = 0.35 x Node Score + 3.64 x AI malignance + 0.00741 x LCRS score + 0.11 x PRS + 0.0229 x Radiomic score

[0122] For non-smokers, the comprehensive score is Multi-dim Score Nonsmoker = 0.48 x Node Score + 3.84 x AI malignance + 0.57 x PRS + 0.0156 x Radiomic score

[0123] wherein Node score is the constructed single modality score; AI malignance is the lung nodule malignancy probability predicted by the deep learning model; LCRS score is the lung cancer epidemiology risk score of the participant, which is calculated by setting the regression coefficient and score value of each variable according to the preset reference group, and adding the score results of all variables to obtain the total LCRS epidemiology risk score, and the specific risk score assignment is shown in Table 2; PRS is the lung cancer polygenic risk score, which is derived from the genotype data of each SNP site, and each SNP site is scored according to wild homozygous type = "0", heterozygous type = "1", and variant homozygous type = "2", and the specific risk score assignment is shown in Table 3; Radiomic score is the imageomics score at the lung lobe level, which includes imageomics features of the left upper lobe (UL), left lower lobe (LL), right upper lobe (UR), right middle lobe (CR) and right lower lobe (LR), and the calculation method of the lung lobe imageomics score is the weighted sum of the product of each imageomics variable and weight, and the specific imageomics score includes variables and weights as shown in Table 4.

[0124] Further, as shown in Figure 3 the present application, the multi-modality score constructed by the present application shows excellent discrimination ability in both smokers and non-smokers, and has obvious advantages compared with the existing mainstream models (Mayo, Brock, VA, PKUPH) and the commonly used 6mm cutoff value (using the average diameter of 6mm as the standard to divide positive nodules, i.e. high-risk population, and non-positive nodules, i.e. low-risk population): in the smoking population, the AUC of the multi-modality score is 0.921, which is significantly higher than that of the Mayo model (0.879), the Brock model (0.894), the VA model (0.860), the PKUPH model (0.784), the 6mm cutoff value (0.794) and the single modality score Node score (0.875); the net reclassification index NRI is 10.28% (95% CI: 5.39%-15.68%), indicating that the comprehensive score has obvious improvement in the reclassification of high-risk individuals.

[0125] In the non-smoking population, the AUC of the multi-modality score is further improved to 0.955, which is also superior to other models (Mayo: 0.833, Brock: 0.894, VA: 0.806, PKUPH: 0.743, 6mm cutoff value: 0.816, single modality score: 0.934); NRI is 7.44% (95% CI: 3.09%-12.85%), showing that the multi-modality score has higher identification efficiency for non-smoking high-risk individuals.

[0126] The multimodal comprehensive score utilizes the integration of the above-mentioned five types of high-value features, combined with a weight optimization algorithm, to effectively improve the identification accuracy of the risk of lung nodules. The comprehensive score shows superior discrimination performance in the data set and has good application prospects, especially for risk stratification, individualized follow-up recommendations and clinical decision support in real lung cancer screening scenarios.

[0127] In summary, the present application can be embedded in the actual lung cancer screening process, integrated into the existing AI screening auxiliary system, and provide single and multimodal scores for clinical practice to improve the identification ability of high-risk populations. In medical centers, public health institutions or lung cancer early screening projects, the model can be deeply integrated with the AI image recognition platform. After the patient completes the low-dose spiral CT (LDCT) or provides biological information, an individualized risk score is automatically generated, and the relative risk level of the screening population is intuitively displayed in the form of a risk percentage, assisting doctors in decision-making and hierarchical intervention.

[0128] In the future, this scheme can be further expanded into an intelligent auxiliary tool for primary screening scenarios, or integrated with image workstations, cloud screening systems and electronic health record platforms, and widely applied to the tumor early screening industry to serve the precise screening and risk management of major chronic diseases such as lung cancer.

[0129] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the above examples do not limit the scope of protection of the present application in any form, and any technical solutions obtained by equivalent replacement or the like fall within the scope of protection of the present application. The parts not covered by the present application are the same as or can be realized by using the prior art.

Claims

1. A multimodal scoring method for differentiating benign and malignant pulmonary nodules, characterized in that, Includes the following steps: Step 1, Data Collection: Collect case information of patients with benign and malignant pulmonary nodules with clear pathological diagnoses, as well as population epidemiological information of participants in the prospective lung cancer screening cohort. Use the two datasets as training and validation sets, respectively. At the same time, collect the SNP detection results of participants in the validation dataset, and collect low-dose spiral CT images of the chest of participants in both datasets. Step 2, Data Preprocessing: Preprocess the collected low-dose spiral CT images; Step 3, Feature Extraction: For the preprocessed low-dose spiral CT images, extract the radiographic features of lung nodules; obtain the deep learning malignancy probability of lung nodules, and obtain the radiomics detection results of the participants' lung lobes; Step 4: Construct and validate a single-modal nodule score: Based on the extracted lung nodule features, construct a single-modal nodule score and evaluate the score's discrimination effect in a validation dataset; Step 5: Construct a multimodal comprehensive score: Based on the single-modal nodule score constructed in Step 4, the participant demographic epidemiological characteristics and SNP detection results collected in Step 1, and the lung nodule deep learning malignancy probability and lung lobe radiomics detection results collected in Step 3, the Cox proportional hazards model and ridge regression model are used for feature screening and weighting to construct a multimodal comprehensive score for lung nodule benign and malignant differentiation.

2. The multimodal scoring method for differentiating benign and malignant pulmonary nodules according to claim 1, characterized in that, In step 1, the data acquisition process includes: The training data came from patients with pulmonary nodules with a clear pathological diagnosis, including both benign and malignant nodules. All cases completed LDCT scans within one month before surgery. Inclusion criteria included: no history of lung surgery, exclusion of other tumor metastases and preoperative intervention. Benign and malignant participants were matched 1:1 by gender and age to construct a certain number of cases as the training set. The validation dataset includes lung cancer tissue screening cohort A and lung cancer opportunistic screening cohort B. Inclusion criteria for cohort A: smokers must have smoked for ≥20 pack-years, including those who previously smoked for ≥20 pack-years and have quit for less than 15 years; non-smokers meeting any of the following criteria were considered high-risk: (1) Living or working in the same room as a smoker who meets the above conditions; a non-smoking woman who has been exposed to secondhand smoke for ≥20 years; (2) A family history of lung cancer in a first-degree relative. (3) Exposure to occupational carcinogens for ≥1 year, (4) History of chronic obstructive pulmonary disease; Cohort B inclusion criteria were: participants aged 18 years or older who were suitable for low-dose spiral CT; collection of participants' demographic epidemiological information; and collection of participants' SNP test results.

3. The multimodal scoring method for differentiating benign and malignant pulmonary nodules according to claim 1, characterized in that, In step 2, the data preprocessing process includes: resampling all low-dose spiral CT images to an isotropic 1×1×1mm. 3 Voxel spacing; image intensity normalization within the standard lung window area; batch effect correction using the ComBat method; automatic segmentation of all lung nodules and the five lobes of the lung.

4. The multimodal scoring method for differentiating benign and malignant pulmonary nodules according to claim 1, characterized in that, In step 3, feature extraction includes: (1) Nodule imaging features: 33 features were extracted, including 18 continuous variables: major axis, minor axis, mean diameter, 3D axis length, volume, mass, surface area, maximum slice area, nodule height, sphericity, compactness, maximum CT value, minimum CT value, average CT value, entropy, skewness, kurtosis, standard deviation, and 15 categorical variables: upper lobe position, middle lobe position, lower lobe position, solid density, ground glass density, partial solid density, spiculation, cavity, spiky protrusion, pleural traction, margin, lobulation, air bronchogram sign, vascular embedding, and vascular clustering. (2) Deep learning malignancy probability: Calculate the malignancy risk probability of nodules using AI malignance; (3) Radiomic features of lung lobes: Based on the lung lobe segmentation results completed in step 2, radiomic features of lung lobes within the region of interest are extracted using the PyRadiomics toolkit.

5. The multimodal scoring method for differentiating benign and malignant pulmonary nodules according to claim 1, characterized in that, In step 4, the process of constructing and validating the single-modality nodule score includes: based on the nodule features extracted in step 3, using the glmnet package in R language to perform LASSO fitting on the nodule features; using ten-fold cross-validation, selecting the optimal regularization parameter λ on the training set; and finally selecting the feature based on lambda.min, which is the λ value with the smallest model cross-validation error, to ensure that the most discriminative variables are retained while controlling the model complexity.

6. The multimodal scoring method for differentiating benign and malignant pulmonary nodules according to claim 5, characterized in that, Based on the selected features and their coefficients fitted by LASSO, a unimodal nodule scoring system is established, and its performance is evaluated on an independent validation set. The specific expression for the unimodal nodule scoring is as follows: Node Score = -18.40994-0.48446×location_l+0.11543×location_c+0.33998×density_ggo -1.32912×density_solid+0.01128×major_diameter-0.00866×minor_diameter-0.0000224×volume-0.000526×hu_max-0.00274×hu_ min+0.00114×hu_mean+0.00096×max_area_mm-0.03487×maximum_3d_diameter+4.82030×sphericity+1.06405×entropy+0.22880×ske wness-0.12753×kurtosis-0.00935×standard_deviation+0.09891×height_diameter-1.85447×lung_spiculation-0.41264×cavitation+5.35330×ratchet_protrusion+1.79332×tractivepleur al+0.40887×margin+0.75031×lung_lobulation+0.65262×air_bronchogram+0.48514×vascular_embedding-1.08487×vascularnotch Wherein, location_l and location_c indicate that the nodule is located in the lower or middle lobe of the lung; density_ggo and density_solid indicate whether the nodule has pure ground-glass density or solid density; major_diameter represents the longest diameter of the nodule measured in cross-section, in millimeters (mm); minor_diameter represents the shortest diameter perpendicular to the longest diameter, in millimeters (mm); and volume represents the total volume occupied by the nodule in three-dimensional space, in mm. 3 hu_max, hu_min, and hu_mean represent the maximum, minimum, and average CT values ​​within the nodule region, respectively, in Hounsfield Units; max_area_mm represents the maximum area of ​​the nodule in a two-dimensional cross-section, in mm. 2 Maximum_3d_diameter represents the longest diameter of the nodule in three-dimensional space, in millimeters (mm); sphericity represents the degree of sphericity of the nodule, with a value closer to 1 indicating a closer approximation to sphericity; entropy is an index of the complexity of grayscale distribution, with a larger value indicating more complex texture; skewness and kurtosis measure the degree of skewness and kurtosis of the grayscale distribution, respectively; standard_deviation represents the standard deviation of the nodule's CT values; height_diameter represents the vertical height of the lung nodule in the longitudinal direction, in millimeters (mm); lung_spiculation represents spiculation; cavitation represents cavitation; ratchet_protrusion represents serrated protrusion; tractivepleural represents pleural traction; margin represents blurred margins; lung_lobulation represents lobulation; air_bronchogram represents air bronchogram; vascular_embedding represents vascular embedding; and vascularnotch represents vascular notching.

7. The multimodal scoring method for differentiating benign and malignant pulmonary nodules according to claim 1, characterized in that, In step 5, the process of constructing a multimodal nodule score includes: on the basis of the single-modal score, further integrating the nodule malignancy probability, population epidemiological characteristics, SNPs, and lung lobe radiomics characteristics output by deep learning; screening predictive factors with univariate association P<0.2 through the Cox proportional hazards model, and using the coefficients output by ridge regression for weighted modeling to establish a multimodal comprehensive discrimination score for benign and malignant lung nodules.

8. The multimodal scoring method for differentiating benign and malignant pulmonary nodules according to claim 7, characterized in that, Independent multimodal composite scores were constructed for smokers and non-smokers respectively. The specific expression of the composite score is as follows: Multi-dim score smoker = 0.35×Node score+3.64×AI malignance+7.41×10 -3 ×LCRS score+0.11×PRS+2.29×10 -2 ×Radiomic score; Multi-dim score nonsmoker = 0.48×Node score+3.84×AI malignance+0.57×PRS+1.56×10 -2 ×Radiomic score; The Node score is the unimodal score constructed in step 4, the AI ​​malignance is the probability of malignancy of lung nodules predicted by the deep learning model, the LCRS score is the participant's lung cancer epidemiological risk score, the PRS is the lung cancer polygenic risk score, and the Radiomic score is the radiomics score at the lobar level.

9. A multimodal scoring method for differentiating benign and malignant pulmonary nodules according to claim 8, characterized in that, The LCRS score is calculated by setting regression coefficients and score values ​​for each variable according to a preset reference group, and then summing the score results of all variables to obtain the total LCRS epidemiological risk score. PRS is a polygenic risk score for lung cancer, derived from the genotype data of each SNP locus extracted in step 3. Each SNP locus is scored according to wild-type homozygosity = 0, heterozygosity = 1, and variant homozygosity = 2. The calculation formula is: PRS = (0.1016 × rs17038564 score) + (0.1174 × rs2293607 score) + (0.1900 × rs11375254 score) + (0.2653 × rs13167280 score) + (0.1447 × rs401681 score) + (0.1826 × rs2517873 score) + (0.0614 × rs3817963 score) + (0.1399 × rs185383 score) (7 rating) + (0.0843 × rs5879422 rating) + (0.1555 × rs4236709 rating) + (0.1004 × rs10429489 rating) + (0.0949 × rs35201538 rating) + (0.0896 × rs4573350 rating) + (0.1338 × rs12265047 rating) + (0.1184 × rs55768116 rating) + (0.0648 × rs11610143 rating) + (0.1002 × rs1200399 rating) + (0.0771 × rs77468143 rating) + (0.1144 × rs200595745 rating); The radiomic score is a radiomics score at the lobar level, namely, UL for the left upper lobe, LL for the left lower lobe, UR for the right upper lobe, CR for the right middle lobe, and LR for the right lower lobe. The radiomics score is calculated by weighted summation of the products of each radiomics variable and its weight.

10. An application of the multimodal scoring construction method as described in any one of claims 1 to 9 in lung nodule risk assessment.

Images