Lung solid nodule identification method based on multi-modal data fusion

Abstract

The embodiment of the invention relates to a lung solid nodule identification method based on multi-modal data fusion. The method comprises the following steps: acquiring multi-modal data of a lung nodule; constructing a training set and a verification set of the multi-modal data; for the training set, the blood sample, the CT image data and the clinical data in the multi-modal data are preprocessed, and methylation features, CT image features and clinical features are obtained in sequence; according to the methylation features, the CT image features and the clinical features, constructing a multi-modal fusion model; according to the verification set, verifying the multi-modal fusion model; and after the verification is passed, obtaining the benign and malignant nodule prediction probability through the multi-modal fusion model.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method for identifying solid lung nodules based on multimodal data fusion. Background Technology

[0002] Lung cancer is the leading cause of cancer-related death worldwide. Its prognosis is closely related to the disease stage at diagnosis. The 5-year overall survival rate for stage IA patients is as high as 85%, while for stage IV patients it is only 6%. Therefore, early diagnosis is crucial for improving cure rates and reducing mortality. Pulmonary nodules, as an early manifestation of lung cancer, require accurate differentiation between benign and malignant nodules for effective clinical decision-making. However, current diagnostic methods have significant limitations. Currently, low-dose computed tomography (LDCT) is the primary method for lung cancer screening, reducing mortality by 20%. However, it is insufficient in differentiating between benign and malignant nodules; 80%-90% of detected nodules are benign, easily leading to overdiagnosis and overtreatment. Furthermore, the results are affected by nodule size, location, and the operator's experience, limiting accuracy. Clinical assessment tools (such as the Mayo Clinic model) rely on imaging parameters and risk factors, but their accuracy is not high, and their sensitivity is greatly affected by nodule characteristics. Invasive diagnostic methods (such as bronchoscopy and transthoracic needle biopsy) can clarify the nature of nodules, but they carry risks of complications such as bleeding and pneumothorax, increasing patient suffering and the medical burden.

[0003] Liquid biopsy, as a non-invasive diagnostic method, has attracted much attention. Among them, circulating cell-free DNA (cfDNA) methylation detection has potential diagnostic value due to its specific changes in the early stages of tumors. However, the accuracy of single biomarker detection (such as traditional tumor markers CEA and SCC) is limited and cannot meet clinical needs alone. When clinical data (such as age and smoking history) or imaging features (such as CT radiomics features) are used alone, they cannot fully characterize the biological characteristics of nodules, both of which carry the risk of misdiagnosis and missed diagnosis. Theoretically, a multimodal fusion strategy integrating cfDNA methylation, CT imaging features, and clinical data can improve diagnostic performance through complementary information. However, current technologies face challenges such as large differences in data structure, difficulties in feature selection, and low efficiency of fusion algorithms, resulting in insufficient model interpretability and clinical applicability. Therefore, developing a technical solution that efficiently integrates multimodal data to overcome the limitations of single methods and achieve accurate differentiation between benign and malignant solid pulmonary nodules has become an urgent clinical problem to be solved. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for identifying solid lung nodules based on multimodal data fusion, in order to address the deficiencies in the existing technology and solve the problems existing in the prior art.

[0005] To achieve the above objectives, this application provides a method for identifying solid lung nodules based on multimodal data fusion, the method comprising: Acquire multimodal data of lung nodules; Construct the training and validation sets for the multimodal data; For the training set, the blood samples, CT image data and clinical data in the multimodal data are preprocessed respectively to obtain methylation features, CT image features and clinical features in sequence; A multimodal fusion model is constructed based on the methylation features, the CT image features, and the clinical features. The multimodal fusion model is validated based on the validation set. Once the verification is successful, the predicted probability of benign or malignant nodules is obtained through the multimodal fusion model.

[0006] In one possible implementation, constructing the training set and validation set of the multimodal data specifically includes: Acquire multimodal data from patients with solid nodules, pure ground-glass nodules, and mixed ground-glass nodules; A first number of multimodal data points and a second number of multimodal data points are obtained from the multimodal data points respectively; the number of solid nodules, pure ground-glass nodules, and mixed ground-glass nodules in the first number of multimodal data points conforms to a first ratio; the number of solid nodules, pure ground-glass nodules, and mixed ground-glass nodules in the second number of multimodal data points conforms to a second ratio; the first number is not equal to the second number; the first ratio is equal to or not equal to the second ratio; Based on the first multimodal data, construct a training set; Based on the second multimodal data, construct a validation set.

[0007] In one possible implementation, obtaining the methylation feature specifically includes: Based on the preset methylation panel, the blood sample to be tested is captured, predicted, and preprocessed to obtain a Bam file; the Bam file records the precise location of each sequencing read aligned to the reference genome, whether the site is methylated, and whether the methylation of the site changes synchronously; Based on preset rules and the Bam file, methylation blocks are divided; the number of cytosine-guanine dinucleotide (CpG) sites in the same methylation block is >2, and the Pearson correlation coefficient of any adjacent CpG sites is >0.5; Based on the Bam file, calculate the number of methylated reads and the total number of reads for each CpG site within each methylation block; The average methylation level (AMF) of each methylated block is obtained by summing the methylated reads of all sites and dividing by the total number of reads of all sites. Extract each individual sequencing sequence from the Bam file; Each sequencing sequence is scored; for a sequencing sequence with k CpG sites, the weight is the first value when all sites are unmethylated, the weight is the second value when all sites are methylated, and the weight is m / k when m consecutive sites are methylated. The methylation haplotype load (MHL) is calculated as MHL = Σ(weight of each sequence) / total number of sequences; where the total number of sequences refers to the total number of sequencing sequences covering the methylation block. Calculate the Pearson correlation coefficient between each AMF or MHL feature and the benign / malignant label, and select the initial methylation features based on the Pearson correlation coefficient. Methylation features were determined from the initial methylation features using the LASSO regression method.

[0008] In one possible implementation, obtaining the Bam file further includes: Blood samples were obtained from multimodal data of lung nodules and blood samples from controls; cfDNA was extracted, sequencing libraries were constructed, and enzymatic transformation was performed to obtain amplified library products; The amplified library products were sequenced using a high-throughput sequencer to output the raw FASTQ file; the FASTQ file includes multiple DNA sequences and their quality information. The original FASTQ file is processed, and fastp is called to add UMI tags to paired reads, filtering out bases that do not meet the quality requirements, and generating a FASTQ file with UMI tags; wherein, filtering out bases that do not meet the quality requirements includes removing the adapter with the cutadapt tool and discarding reads with a length <50bp; the first 12 bases of Read2 are UMI tags; The BisMark tool was used to align the filtered paired reads to the hg19 human reference genome to generate the initial Bam file. Based on the UMI tag, repetitive sequences introduced by PCR amplification were removed. The initial Bam file was sorted by chromosome position using Samtools and reads with an alignment quality >20 were selected. Reads with a CT conversion rate <95% at non-CpG sites were filtered to generate the Bam file and create an index.

[0009] In one possible implementation, obtaining the CT image features specifically includes: Obtain the first region of interest (ROI) and the second region of interest (ROI) from the CT image data determined by the first terminal and the second terminal, respectively. The consistency between the first ROI and the second ROI was verified using the Kappa coefficient; Once the verification is successful, the target ROI is determined. Extract the radiomics features of the target ROI; Calculate the intraclass correlation coefficient (ICC) value of the radiomics features; The initial radiomics features corresponding to ICC values ​​greater than the first threshold are retained; The variance of each initial radiomics feature is calculated, and the variance is compared with a preset second threshold. The initial CT image features with variances greater than the second threshold are retained as CT image features.

[0010] In one possible implementation, the CT image features include morphological features, grayscale features, texture features, and higher-order features; the morphological features describe the geometry of the nodules, the grayscale features describe the distribution of pixel values ​​within the ROI, the texture features describe the spatial distribution and relationships of pixel grayscale values, the relationships including contrast and correlation, and the higher-order features are features extracted after filtering the image.

[0011] In one possible implementation, obtaining the clinical characteristics specifically includes: Clinical data are constructed based on demographic information, medical history information, tumor markers, hematological parameters, and derived indicators. Each variable in the clinical data is processed to determine the difference of each variable between the benign and malignant groups; the processing includes using the T test or Mann-Whitney U test for continuous variables, the chi-square test for categorical variables, and the rank-sum test for ordinal variables; Variables with p-values ​​less than the third threshold are retained to obtain the first variable; For the first variable, the second variable is obtained by screening using LASSO regression and Logistic regression with minimum absolute contraction and selection operators. The second variable is binarized according to the critical value to obtain a binary classification feature, which is a clinical feature.

[0012] In one possible implementation, constructing the multimodal fusion model based on the methylation features, the CT image features, and the clinical features specifically includes: Z-score normalization was applied to methylation features, normalization was applied to clinical features, and L2 regularization was applied to CT image features. The SHAP values ​​of processed methylation features, CT imaging features, and clinical features were evaluated, and the features were sorted according to their SHAP values, with the top n features selected. The first n features are decomposed using a low-rank matrix to obtain the decomposition results; Element-wise product fusion is performed using F=∏(M_i×W_i) to fuse all mode decomposition results; where ∏ is the element-wise product operation, M_i is the feature matrix of the i-th mode, and W_i is the mode weight matrix. Logistic regression is used as the base model, support vector machine is used as the meta-model, and they are stacked and integrated. The multimodal fusion model is then optimized using the cross-entropy loss function to obtain the multimodal fusion model.

[0013] In one possible implementation, the method further includes: The overall performance of the multimodal fusion model was verified using an independent test set; wherein the overall performance was measured by the area under the ROC curve (AUC), sensitivity, specificity, and accuracy.

[0014] In one possible implementation, the method further includes: Based on the methylation features and validation set, the CT image features and validation set, and the clinical features and validation set, respectively, a methylation feature model, a CT image feature model, and a clinical feature model are constructed: The training process involves using multiple methylation features of each sample in the training set and the labels corresponding to those methylation features as input to train a methylation feature model. The performance of the methylation feature model is evaluated using 5-fold cross-validation, and the parameters are adjusted until optimal. The methylation feature model is then evaluated using a test set, and the training ends when AUC ≥ 0.9. The specific steps of constructing a CT image feature model based on CT image features are as follows: taking multiple CT image features of each sample in the training set and the labels corresponding to the CT image features as input, training the CT image feature model, evaluating the performance of the CT image feature model using 5-fold cross-validation, adjusting the parameters until optimal, evaluating the CT image feature model through the test set, and ensuring that AUC ≥ 0.9, at which point the training of the CT image feature model is complete. The specific steps of constructing a clinical feature model based on clinical characteristics include: taking multiple clinical features of each sample in the training set and the labels corresponding to the clinical features as input, training the clinical feature model, evaluating the performance of the clinical feature model using 5-fold cross-validation, adjusting the parameters until optimal, evaluating the clinical feature model through the test set, and completing the training of the clinical feature model when AUC ≥ 0.9. Based on the outputs of the methylation feature model, CT image feature model, and clinical feature model, a multimodal fusion model is constructed.

[0015] The method for identifying solid lung nodules based on multimodal data fusion in this application combines methylation features, CT image features, and other features to construct a multimodal fusion model for training and validation. Through validation on the test set, the model can output the predicted probability of benign and malignant nodules, thereby greatly improving the prediction accuracy. Attached Figure Description

[0016] Figure 1 This is one of the flowcharts of the method for identifying solid lung nodules based on multimodal data fusion provided in the embodiments of this application; Figure 2 This is a schematic diagram of the final sample; Figure 3 for Figure 1 A flowchart of methylation characteristics was obtained. Figure 4 for Figure 3 The flowchart of the Bam file is obtained from the process. Figure 5 for Figure 1 The flowchart for obtaining CT image features is as follows; Figure 6 for Figure 1 The clinical feature flowchart was obtained from the process; Figure 7 for Figure 1 Flowchart for step 140; Figure 8 The second flowchart of the method for identifying solid lung nodules based on multimodal data fusion provided in this application; Figure 9 The third flowchart of the method for identifying solid lung nodules based on multimodal data fusion provided in this application embodiment; Figure 10 A comparison chart of ROC curves for single-modal and multi-modal fusion in this application for training set; Figure 11 This is a comparison chart of ROC curves for single-modal and multi-modal fusion in this application, used as a test set. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0019] Figure 1 The flowchart of the method for identifying solid lung nodules based on multimodal data fusion provided in the embodiments of the present invention is as follows: Figure 1 As shown, the method includes the following steps: Step 110: Obtain multimodal data of lung nodules.

[0020] Step 120: Construct the training set and validation set of the multimodal data.

[0021] The process involves acquiring multimodal data from patients with solid nodules, pure ground-glass nodules, and mixed ground-glass nodules; acquiring a first number of multimodal data points and a second number of multimodal data points from the multimodal data points; ensuring that the number of solid nodules, pure ground-glass nodules, and mixed ground-glass nodules in the first number of multimodal data points conforms to a first ratio; ensuring that the number of solid nodules, pure ground-glass nodules, and mixed ground-glass nodules in the second number of multimodal data points conforms to a second ratio; wherein the first number is not equal to the second number; and the first ratio is equal to or not equal to the second ratio; constructing a training set based on the first multimodal data points; and constructing a validation set based on the second multimodal data points.

[0022] Specifically, samples were collected from patients with pulmonary nodules admitted to the thoracic surgery, respiratory, and radiology departments of the hospital. A final selection of 3000 eligible samples was completed, covering common clinical pulmonary nodule subtypes. The following inclusion criteria and imaging characteristics of each subtype were required to be met simultaneously: Basic inclusion criteria: age ≥ 18 years, gender not limited, all patients signed informed consent forms, and this study protocol was approved by the ethics committees of the participating centers; Common characteristics and subtypes of nodules: Lung nodules diagnosed by low-dose computed tomography (LDCT) or conventional CT (excluding multiple nodules with a diameter ≥5mm), with a diameter range of 5-30mm, are identified. Specific subtypes and imaging characteristics are as follows: ① Solid nodules (1800 cases): These nodules have a higher density than lung tissue in CT images, with clear or blurred boundaries, and can completely obscure pulmonary blood vessels and bronchial structures. ②600 cases of pure ground-glass nodules: The density of the nodules was slightly increased in CT images, appearing as a cloud-like structure, without obscuring the pulmonary blood vessels and bronchial structures; ③ 600 cases of mixed ground-glass nodules: CT images contained both ground-glass density components and solid density components, with the solid component accounting for 10%-90%.

[0023] Collect blood samples, CT imaging data, and clinical data from patients meeting the above criteria; these three are collectively referred to as multimodal data. Patients with benign or malignant nodules must meet the following confirmation criteria (applicable to all subtypes): Benign nodules: confirmed by postoperative pathology or comprehensive expert evaluation, with no change observed during follow-up for ≥12 months; Malignant nodules: confirmed by postoperative pathology. Ultimately, multimodal data from 3000 patients with pulmonary nodules conforming to the above types (including 3 subtypes) were included for the study. The 3000 samples were divided into a training set and an independent validation set in a 7:3 ratio: 2100 cases in the training set (with cases from each subtype allocated in a 6:2:2 ratio to ensure balanced subtype distribution) and 900 cases in the independent validation set (with cases from each subtype allocated in a 6:2:2 ratio). This ensured multi-center data, with at least 5 clinical centers and balanced baseline characteristics. If the malignancy rate of each subtype remained similar between the training and validation sets, the original baseline feature matching logic was used for model training and validation.

[0024] Clinical data included tumor markers and other hematological parameters, specifically covering CA125 (ng / ml), SCC (ng / ml), CEA (ng / ml), lymphocyte count, platelet count, lactate dehydrogenase (LDH) (U / L), NLR, cytokeratin 19 fragment (ng / ml), monocytes, neuron-specific enolase (ng / ml), progastrin-releasing peptide (progastrin-releasing peptide) (ng / ml), neutrophils, LMR, PLR, as well as sex, age, and smoking history. Final sample data were as follows: Figure 2 As shown.

[0025] Step 130: For the training set, preprocess the blood samples, CT image data and clinical data in the multimodal data respectively to obtain methylation features, CT image features and clinical features in sequence.

[0026] The following sections explain how to obtain methylation features, CT imaging features, and clinical features.

[0027] First, methylation characteristics like Figure 3 As shown in the figure, the methylation characteristics are obtained, including the following steps: Step 310: According to the preset methylation panel, capture prediction and preprocessing are performed on the blood sample to be detected to obtain the Bam file; the Bam file records the precise position of each sequencing read aligned to the reference genome, whether the site is methylated, and whether the methylation of the site changes synchronously. The Bam file serves as a bridge between the raw sequencing data and methylation features. All determinations of methylation blocks and calculations of their features are based on the extraction, processing, and interpretation of the information contained in the Bam file.

[0028] Step 320: Based on preset rules and the Bam file, divide the methylation blocks; the number of cytosine-guanine dinucleotide (CpG) sites in the same methylation block is >2, and the Pearson correlation coefficient of any adjacent CpG sites is >0.5; Step 330: Based on the Bam file, calculate the number of methylated reads and the total number of reads for each CpG site within each methylation block; Step 340: Sum the methylated reads of all sites and divide by the total number of reads of all sites to obtain the average methylation level (AMF) of each methylated block; Step 350: Extract each individual sequencing sequence from the Bam file; Step 360: Score each sequencing sequence; for a sequencing sequence with k CpG sites, the weight is the first value when all sites are unmethylated, the weight is the second value when all sites are methylated, and the weight is m / k when m consecutive sites are methylated. The first value is 0, and the second value is 1.

[0029] Step 370: Calculate the methylation haplotype load (MHL) according to MHL = Σ(weight of each sequence) / total number of sequences; where the total number of sequences refers to the total number of sequencing sequences covering the methylation block. Step 380: Calculate the Pearson correlation coefficient between each AMF or MHL feature and the benign / malignant label, and screen the initial methylation features based on the Pearson correlation coefficient; Specifically, the AMF or MHL of each MB is checked for significant differences between the benign and malignant groups. If the Pearson correlation coefficient is greater than a preset threshold, the corresponding methylation feature can be retained.

[0030] Step 390: Determine the methylation features from the initial methylation features using the LASSO regression method.

[0031] Specifically, the salient features obtained in the previous step are input into the LASSO model, and the coefficients of unimportant and redundant features are compressed to 0, thereby automatically selecting the most discriminative and non-redundant features.

[0032] Furthermore, such as Figure 4 As shown, the process of obtaining the Bam file also includes the following steps: Step 410: Obtain blood samples from the multimodal data of lung nodules and blood samples from controls; extract cfDNA, construct sequencing libraries, and perform enzymatic transformation to obtain amplified library products; Step 420: Sequencing the amplified library product using a high-throughput sequencer to output the original FASTQ file; the FASTQ file includes multiple DNA sequences and their quality information; Step 430: Process the original FASTQ file and call fastp to add UMI tags to paired reads, filter out bases that do not meet the quality requirements, and generate a FASTQ file with UMI tags; wherein, filtering out bases that do not meet the quality requirements includes removing the adapter with the cutadapt tool and discarding reads with a length <50bp; the first 12 bases of Read2 are UMI tags; Step 440: Use the BisMark tool to align the filtered paired reads to the hg19 human reference genome to generate the initial Bam file. Based on the UMI tag, remove repetitive sequences introduced by PCR amplification. Use Samtools to sort the initial Bam file by chromosome position and screen reads with an alignment quality >20. Filter reads with a CT conversion rate <95% at non-CpG sites to generate the Bam file and create an index.

[0033] In a specific implementation, the method for obtaining methylation features is explained. To address the challenge of differentiating between benign and malignant pulmonary nodules, this application first provides a self-developed multi-gene methylation panel. By analyzing the methylation degree and haplotype diversity of linked regions within the panel, and combining this with a machine learning model, accurate identification of benign and malignant pulmonary nodules is achieved. Specific methods include: (1) Provides a multi-gene combination for detecting methylation; (2) The plasma sample to be tested was captured, sequenced and preprocessed using the methylated panel described above to obtain the Bam file; (3) Methylation Haplotype Blocks (MBs) are divided based on preset rules: the number of CpG sites in the same MB is >2, and the Pearson correlation coefficient between any two adjacent CpG sites is >0.5; the average methylation frequency (AMF) of each MB is calculated, which is the ratio of the sum of the number of methylated reads of all CpG sites in the region to the total number of reads (methylated + unmethylated) in the region; at the same time, the methylation haplotype load (MHL) of each MB is calculated, and the block length and the proportion of continuous methylated sites are considered comprehensively to evaluate the diversity of methylation haplotypes. (4) Compare the detected AMF with the normal AMF level of the corresponding gene in the population, and / or compare the detected MHL with the corresponding MHL level in the population to identify the benign or malignant status of the lung nodules.

[0034] In some implementations, the method further includes performing the above tests again after the subject has received the intervention, and monitoring changes in the status of lung nodules (such as the stability of benign nodules or the progression / remission of malignant nodules) by comparing the two AMF and / or MHL results.

[0035] In some implementations, step (1) above includes extracting DNA from biological samples (such as blood, serum, plasma, sputum, pleural fluid, bronchoalveolar lavage fluid, etc.), treating it with bisulfite (to deaminate unmethylated cytosine while keeping methylated cytosine unchanged), and then performing capture sequencing and preprocessing through the above-mentioned methylation panel to obtain a Bam file, and dividing it into MBs according to the above-mentioned rules.

[0036] In some implementations, the methylation-specific primer pair used for multi-gene AMF detection in step (1) includes primers adapted to the multi-gene combination (such as SEQ ID NO: 1-XXX, which are adjusted according to the number of genes), and the AMF of each MB needs to be detected separately.

[0037] In some implementation schemes, the identification of benign or malignant pulmonary nodules can be achieved by: logistic regression analysis based on AMF alone, logistic regression analysis based on MHL alone, or logistic regression analysis combining AMF and MHL.

[0038] In another specific implementation, this application provides a kit for differentiating benign from malignant pulmonary nodules, comprising reagents for detecting AMF and MHL of the aforementioned multi-gene methylation blocks (MB) in biological samples, specifically including methylation-specific primer pairs (adapted to multi-gene combinations) and optional bisulfite for the amplification reaction. The kit also includes instructions for use, outlining the detection procedure and a method for differentiating benign from malignant pulmonary nodules by processing the results using logistic regression.

[0039] This invention also provides a next-generation sequencing detection system based on methylation multiplex amplification, comprising: (1) Sample processing module: Obtain plasma samples from patients with pulmonary nodules and controls, extract cfDNA, construct a library and perform enzymatic transformation to obtain amplified library products; (2) Sequencing module: Performs methylation sequencing on the amplified library products; (3) Data filtering module: Process the original FASTQ file, call fastp to add UMI tags to paired reads (the first 12 bases of Read2 are UMI) and filter low-quality bases to generate FASTQ file with UMI; use cutadapt to remove the adapter and discard reads with a length <50bp. (4) Alignment module: BisMark is called to align paired reads to the hg19 human reference genome to generate the initial Bam file; based on UMI to remove duplicate sequences (deduplicate_bismark), Samtools is used to sort by chromosome position and filter reads with alignment quality >20, and non-CpG reads with CT conversion rate <95% are filtered to generate the final Bam file and build an index (Samtools index); (5) AMF extraction module: Based on the alignment results, obtain the number of methylated and unmethylated reads at each CpG site in the whole genome, and calculate the AMF of each MB; (6) MHL extraction module: Calculates MHL based on MB length and the proportion of continuous methylation sites; AMF and MHL were screened using Pearson correlation coefficient and LASSO method to obtain features that meet the threshold for the differentiation of benign and malignant pulmonary nodules.

[0040] Second, CT imaging features like Figure 5 As shown, obtaining CT image features includes the following steps: Step 510: Obtain the first region of interest (ROI) and the second region of interest (ROI) in the CT image data determined by the first terminal and the second terminal, respectively. The first terminal and the second terminal are interrupts used by different doctors.

[0041] Step 520: Verify the consistency between the first ROI and the second ROI using the Kappa coefficient; Step 530: Once the verification is successful, determine the target ROI; Step 540: Extract the radiomics features of the target ROI; Step 550: Calculate the intragroup correlation coefficient (ICC) value of the radiomics features; Step 560: Retain the initial CT image features corresponding to ICC values ​​greater than the first threshold; Step 570: Calculate the variance of each initial CT image feature, compare the variance with a preset second threshold, and retain the initial CT image features whose variance is greater than the second threshold to obtain the CT image features.

[0042] The CT image features include morphological features, grayscale features, texture features, and higher-order features. The morphological features describe the geometry of the nodules, the grayscale features describe the distribution of pixel values ​​within the ROI, the texture features describe the spatial distribution and relationships of pixel grayscale values, including contrast and correlation, and the higher-order features are features extracted after filtering the image.

[0043] In a specific implementation, image data acquisition is the foundation of radiomics analysis, and standardized equipment, parameters, and annotation processes are required to ensure the consistency and accuracy of the data.

[0044] Imaging Equipment and Parameter Settings: Chest images were acquired using a 128-slice CT scanner with a slice thickness set to 1.25mm to reduce inter-slice artifacts and ensure clarity of nodule details. Image data was stored in DICOM format, a universal medical imaging format containing metadata such as patient information and scan parameters. Simultaneously, 3D reconstruction data in NRRD format was generated for subsequent 3D feature analysis. During scanning, window width and level were standardized; for example, for the lung window: WW=1500, WL=-600, to avoid image grayscale deviations caused by parameter differences. Region of Interest (ROI) Marking: Two radiologists with ≥5 years of experience independently used MITK software, an open-source medical image processing tool, to manually segment the lung nodule region and determine the ROI, i.e., the physical boundary of the nodule. The first terminal was the terminal used by the first physician, and the second terminal was the terminal used by the second physician. During the annotation process, consistency must be verified by the Kappa coefficient, which must be ≥0.85. If there are any discrepancies in the annotation, a third associate chief physician will arbitrate to ensure the accuracy of the ROI boundary (to reduce the impact of human error on subsequent feature extraction).

[0045] Image feature extraction: Feature extraction is the core of radiomics. It requires the use of specialized tools to extract quantifiable image features from the ROI, covering multiple dimensions to comprehensively characterize the biological characteristics of nodules. Extraction Tools and Feature Types: Pyradiomics, an open-source radiomics feature extraction library, was used to extract features from DICOM and NRRD 3D reconstruction data. Initially, 1800 radiomics features were obtained, covering the following core dimensions: Morphological features: describing the geometric shape of nodules, such as volume, surface area, sphericity, longest diameter, edge smoothness, etc., reflecting the physical morphological differences of nodules; Gray-level features: statistical features based on the gray-level distribution of image pixels, such as mean, standard deviation, skewness, kurtosis, etc., reflecting the density uniformity within nodules; Texture features: describing the spatial distribution pattern of pixel gray levels, such as contrast, correlation, and entropy based on the gray-level co-occurrence matrix (GLCM), and long run advantage based on the gray-level run matrix (GLRLM), reflecting the heterogeneity of the internal structure of nodules; Higher-order features: features extracted after processing by wavelet transform, Laplace filtering, etc., enhancing the capture of subtle structural differences.

[0046] Feature Selection (Stability Verification): Since the initially extracted features may contain redundant or unstable features, and are affected by annotation bias and scanning noise, stable features need to be selected through statistical methods: Calculating Intragroup Correlation Coefficient (ICC): Features are extracted separately for the same batch of nodules annotated by two physicians, and the ICC values ​​of the two extraction results are calculated. Features with ICC ≥ 0.8 are retained, i.e., features with a first threshold ≥ 0.8, indicating high consistency and strong stability among different annotators. Removing Low-Variance Features: Features with extremely small value variations in all samples, such as standard deviations close to 0, are excluded to avoid features with no discriminative power interfering with the model. After selection, stable features are finally retained as input for subsequent model construction. The second threshold can be set to a positive number close to 0, for example, variance less than 0.001 or standard deviation less than a certain minimum value. In a specific example, this could be "the second threshold can be 0", or "features with variance less than 0.001 are excluded".

[0047] Third, clinical features like Figure 6 As shown, obtaining clinical characteristics includes the following steps: Step 610: Based on demographic information, medical history information, tumor markers, hematological parameters, and derived indicators, clinical data are constructed; The demographic information includes age and sex. Medical history includes smoking history (with calculated "pack-years"). Tumor markers include CEA, SCC, CA125, and cytokeratin 19 fragments. Hematological parameters include lymphocyte count, neutrophil count, and platelet count. Derivative indicators include NLR (neutrophil-to-lymphocyte ratio) and PLR (platelet-to-lymphocyte ratio).

[0048] Next, data cleaning is performed, including handling missing values. For variables with few missing values, the mean / median can be used to fill in missing samples or to delete missing samples; if there are too many missing values, the variable should be removed. Outliers are handled: obvious data entry errors or physiologically impossible extreme values ​​are checked and addressed. Data integration: ensure that the data comes from standardized testing methods across centers, and perform corrections as necessary.

[0049] Step 620: Process each variable in the clinical data to determine the difference of each variable between the benign and malignant groups; the processing includes using the T test or Mann-Whitney U test for continuous variables, the chi-square test for categorical variables, and the rank-sum test for ordinal variables. Specifically, each clinical variable is statistically analyzed against the "gold standard" (pathological outcome: benign = 0, malignant = 1). For continuous variables, such as age and CEA value, a t-test or Mann-Whitney U test is used (chosen based on whether the data conforms to a normal distribution) to compare their differences between the benign and malignant groups. For categorical variables, such as gender, a chi-square test is used. For ordinal variables, such as smoking history grouping, a rank-sum test is used.

[0050] Step 630: Retain the variables whose P-values ​​are less than the third threshold to obtain the first variable; Specifically, the third threshold can be 0.05, retaining variables with a p-value < 0.05. This indicates that the difference in this variable between the benign and malignant groups is not caused by random error and is statistically significant. For example, from 30 initial variables, 10 variables that are significantly related to benign or malignant can be selected and called the first variables.

[0051] Step 640: For the first variable, the second variable is obtained by screening through LASSO regression and Logistic regression with minimum absolute shrinkage and selection operators. Specifically, all the selected primary variables are simultaneously added to the Logistic regression model, usually using the "backward" or "forward" method. Based on the P-value, if P>0.05, the variable is removed. The least significant variables are gradually removed until all variables in the Logistic regression model are significant.

[0052] Step 650: The second variable is binarized according to the critical value to obtain a binary classification feature, which is a clinical feature.

[0053] Specifically, for continuous variables selected by the LASSO regression model, such as age, CEA, and NLR, it is necessary to determine the optimal cutoff value for distinguishing between benign and malignant diseases through ROC curve analysis or according to clinically recognized standards.

[0054] For example: age ≥ 60 years, smoking history ≥ 20 pack-years, CEA ≥ 5 ng / mL, SCC ≥ 1.5 ng / mL, NLR ≥ 3, platelet count ≥ 300 × 10⁻⁶. 9 / L.

[0055] Convert continuous variables into binary variables (0 or 1) based on critical values. For example, convert the continuous variable "age" into "whether the age is ≥60 years old", where yes = 1 and no = 0.

[0056] Step 140: Construct a multimodal fusion model based on the methylation features, the CT image features, and the clinical features.

[0057] Specifically, such as Figure 7 As shown, step 140 includes the following: Step 1401: Perform Z-score standardization on methylation features, normalize clinical features, and perform L2 regularization on CT image features. Methylation features were standardized using Z-scores, clinical features were normalized using min-max normalization (mapped to the [0,1] interval), and imaging features were processed using L2 regularization. The three types of features were then standardized again. Finally, all modal features were concatenated, for example, 120+79+6 features were concatenated sequentially to obtain the feature matrix.

[0058] Step 1402: Use SHAP values ​​to evaluate the SHAP values ​​of processed methylation features, CT image features and clinical features, and sort them according to SHAP values, selecting the top n features in the sort. Specifically, the SHAP value is calculated using the base model, and then the mean of the SHAP values ​​is calculated. Based on the SHAP values ​​and the mean of the SHAP values, the top n key fusion features are selected and ranked. Here, n can be set to an empirical value of 130.

[0059] Step 1403: Perform low-rank decomposition on the first n features using a low-rank matrix to obtain the decomposition results; Specifically, truncated SVD is used for low-rank decomposition, the explained variance ratio is calculated, and then low-rank decomposition is performed.

[0060] Step 1404: Element-wise product fusion is performed using F=∏(M_i×W_i) to fuse all mode decomposition results; where ∏ is the element-wise product operation, M_i is the feature matrix of the i-th mode, and W_i is the mode weight matrix. Specifically, low-rank matrix decomposition is used to achieve dimensionality unification, and all modal decomposition results are fused through element-wise product, where the rank parameter r can be 30.

[0061] Step 1405: Logistic regression is used as the base model, support vector machine is used as the meta-model, and they are stacked and integrated. The multimodal fusion model is then optimized using the cross-entropy loss function to obtain the multimodal fusion model.

[0062] Specifically, we define a base model (first layer) and a meta-model (second layer), use StackingClassifier to combine the base model and meta-model into a stacked model, and then train the stacked model using the processed features and labels.

[0063] The base model is Logistic Regression (LR), the meta-model is Support Vector Machine (SVM), and the model is optimized using the cross-entropy loss function to output the predicted probability (0-100%) of nodules as benign or malignant.

[0064] Step 150: Validate the constructed multimodal fusion model according to the validation set.

[0065] Specifically, the model performance was verified using an independent test set, and the identification effect was evaluated using ROC curves, AUC values, sensitivity, and specificity.

[0066] The multimodal joint model has an AUC ≥ 0.95, sensitivity ≥ 85%, and specificity ≥ 95% on the test set, with sensitivity ≥ 87.7% for 5-10 mm nodules, sensitivity ≥ 93% for 11-20 mm nodules, and sensitivity ≥ 95% for 21-30 mm nodules.

[0067] Step 160: After verification, the predicted probability of benign or malignant nodules is obtained through the multimodal fusion model.

[0068] Specifically, once the model is built, the probability of benign or malignant nodules can be predicted by inputting the current multimodal data.

[0069] Furthermore, in an optional implementation, this application includes a step of constructing three unimodal models before constructing the multimodal fusion model, and then fusing the results of the three unimodal models to construct the multimodal fusion model. See also Figure 8 The specific steps are as follows: Step 810: Construct a methylation feature model, a CT image feature model, and a clinical feature model sequentially based on the methylation features and validation set, the CT image features and validation set, and the clinical feature and validation set. The training process involves using multiple methylation features of each sample in the training set and the labels corresponding to those methylation features as input to train a methylation feature model. The performance of the methylation feature model is evaluated using 5-fold cross-validation, and the parameters are adjusted until optimal. The methylation feature model is then evaluated using a test set, and the training ends when AUC ≥ 0.9. The algorithm chosen is Logistic Regression (LR). Input features include 120 methylation-related features. Optimization parameters include L2 regularization as the penalty term, regularization strength C=0.1, and liblinear as the solver (determined through 5-fold cross-validation). Performance requirement: AUC ≥ 0.90 on the test set. For example, for the 2100 samples in the training set, each sample has 120 methylation features, corresponding to binary labels: nausea = 1, benign = 0. Therefore, this methylation feature model can output predicted probabilities.

[0070] The construction of a CT image feature model based on CT image features specifically includes: taking multiple CT image features of each sample in the training set and the labels corresponding to the CT image features as input, training the CT image feature model, evaluating the performance of the CT image feature model using 5-fold cross-validation, adjusting the parameters until optimal, evaluating the CT image feature model through the test set, and completing the training of the CT image feature model when AUC ≥ 0.75. Algorithm selection: Logistic regression (LR), input features: 79 CT image-derived features (including morphological, texture and other radiomic features), optimization parameters: penalty term is L1 regularization, regularization strength C=0.05, solver is saga (determined through 5-fold cross-validation), performance requirement: test set AUC≥0.75.

[0071] The specific steps of constructing a clinical feature model based on clinical characteristics include: taking multiple clinical features of each sample in the training set and the labels corresponding to the clinical features as input, training the clinical feature model, evaluating the performance of the clinical feature model using 5-fold cross-validation, adjusting the parameters until optimal, evaluating the clinical feature model through the test set, and completing the training of the clinical feature model when AUC ≥ 0.7. Algorithm selection: Logistic regression (LR), input features: 6 key clinical features (such as age, medical history, laboratory indicators, etc.), optimization parameters: L2 regularization as the penalty term, lbfgs as the solver, regularization strength C=1.0 (determined through 5-fold cross-validation), performance requirement: AUC ≥ 0.70 on the test set.

[0072] Step 820: Based on the outputs of the methylation feature model, CT image feature model, and clinical feature model, a multimodal fusion model is constructed.

[0073] A multimodal model can be constructed based on the predicted probabilities output by each unimodal model.

[0074] In one possible implementation, see Figure 9 The blood samples used in this embodiment were obtained from patients with pulmonary nodules admitted to the thoracic surgery, respiratory and radiology departments of a prospective hospital. The experimental protocol has been approved by the institution's ethics committee, and all patients have signed informed consent forms.

[0075] Experimental reagents and instruments included: cell-free DNA preservation tubes, such as Kangwei Century, catalog number CW2816M; QIAampMinElute circulating cell-free DNA micro-extraction kit (Qiagen, catalog number 55204); EpiTect rapid DNA bisulfite conversion kit (Qiagen, catalog number 59824); QIAseq targeted methylation detection panel (containing 579 lung cancer-specific methylation regions); Illumina NextSeq 500 sequencing platform; 128-slice CT scanner; MITK image segmentation software; and Pyradiomics feature extraction tool.

[0076] Sample recruitment and grouping 1. Sample inclusion criteria This embodiment adopted a prospective enrollment method. The samples were prospectively collected from patients with pulmonary nodules treated in the thoracic surgery, respiratory, and radiology departments of a hospital. A total of 3,000 eligible samples were ultimately selected. The specific inclusion criteria are as follows: Basic requirements: age ≥ 18 years, gender not limited, all patients signed informed consent forms, and this study protocol was approved by the ethics committees of the participating centers; Nodule characteristics: Pulmonary nodules were diagnosed by low-dose computed tomography (LDCT) or conventional CT scans. The nodule diameter ranged from 5 to 30 mm, and the nodule types covered common clinical subtypes: ① Solid nodules (1800 cases, with density higher than lung tissue in CT images, clear or blurred borders, which can completely obscure pulmonary vessels and bronchial structures); ② Pure ground-glass nodules (600 cases, with slightly increased density in CT images, appearing cloudy, and not obscuring pulmonary vessels and bronchial structures); ③ Mixed ground-glass nodules (600 cases, with both ground-glass and solid density components in CT images, with the solid component accounting for 10%-90%). Clinical status: Exclude pregnant women and breastfeeding women; exclude patients with signs of metastasis such as pleural effusion and mediastinal lymph node short axis diameter >10mm; exclude patients diagnosed with malignant tumors other than melanoma or skin cancer within the past 2 years; exclude patients who have not undergone subsequent confirmation of benign or malignant nature (pathological examination or long-term follow-up).

[0077] 2. Criteria for confirming the benignity or malignancy of samples To ensure the accuracy of benign and malignant labeling of samples, a dual confirmation method of "gold standard + long-term follow-up" is adopted, as follows: Malignant nodule confirmation: All malignant nodule samples were confirmed by pathological examination. The pathological types included non-small cell lung cancer (squamous cell carcinoma, adenocarcinoma, large cell carcinoma) and small cell lung cancer. The pathological methods included: ① postoperative pathological examination after surgical resection of nodule tissue; ② histopathological examination of tissue obtained by CT-guided transthoracic needle biopsy and bronchoscopic biopsy; ③ histopathological examination of tissue obtained by thoracoscopic surgery. Confirmation of benign nodules: Two confirmation pathways are used, and inclusion is possible if either one is met: ① Pathological confirmation: Tissue is obtained through the above-mentioned biopsy or surgical methods, and the pathological diagnosis is benign lesion (including inflammatory nodules, tuberculomas, hamartomas, lipomas, fibromas, etc.); ② Follow-up confirmation: No invasive pathological examination is performed, but regular follow-up is conducted via LDCT (every 3-6 months, with a cumulative follow-up time of ≥12 months), and there are no significant changes in nodule size, shape, or density, and no new lesions appear. The nodule is jointly determined to be benign by two physicians with ≥5 years of experience in chest imaging diagnosis.

[0078] 3. Sample grouping and baseline characteristics Grouping method: The 3,000 qualified samples were divided into a training set and an independent validation set using a stratified random sampling method at a ratio of 7:3. This ensured that the two sets of samples were balanced and consistent in baseline characteristics such as nodule type, benign / malignant ratio, age distribution, and gender ratio, thus avoiding sampling bias: Training set: 2,100 cases; Independent validation set: 900 cases.

[0079] Baseline feature matching: There were no statistically significant differences in baseline characteristics between the two groups (P > 0.05), specifically as follows: age distribution (training set: 45-78 years, mean 62.3 ± 8.5 years; validation set: 46-77 years, mean 61.8 ± 8.2 years); gender ratio (training set: male:female = 1.2:1; validation set: male:female = 1.1:1); percentage of smokers with a history of smoking (45.2% of smokers had a history of smoking ≥ 20 pack-years in the training set; 44.7% in the validation set); and mean nodule diameter (training set: 15.6 ± 6.3 mm; validation set: 15.2 ± 6.1 mm).

[0080] 4. Sample quality control To ensure the effectiveness of subsequent multimodal data extraction, the enrolled samples underwent rigorous quality screening, resulting in the removal of 52 unqualified samples (ultimately retaining 3000 samples). The quality control standards are as follows: Plasma sample quality: 5 mL of peripheral venous blood was collected from the patient using a cell-free DNA preservation tube (Kangwei Century, CW2816M). After separating the plasma, circulating cell-free DNA (cfDNA) was extracted. The concentration of cfDNA was detected by Qubit fluorescence quantitative method. The concentration was required to be ≥1 ng / μL, and the DNA fragment integrity met the requirements of targeted methylation sequencing (fragment length was mainly distributed in 160-180 bp). CT image quality: The CT imaging equipment is a 128-slice or higher resolution CT scanner. The image data is stored in DICOM format. The slice thickness is required to be ≤1.25mm. The window width and window level are uniformly set to lung window (WW=1500HU, WL=-600HU). The images are free of motion artifacts, metal artifacts and respiratory artifacts. The nodule area is clearly distinguishable. Clinical data quality: Clinical data is collected through structured forms, including demographic information (age, gender, smoking history), tumor markers (CEA, SCC, CA125, etc.), and hematological parameters (lymphocyte count, NLR ratio, etc.). Data completeness is required to be ≥95%, with no missing key information, and all laboratory test indicators are issued by the certified laboratories of each center, and the test methods comply with national standards.

[0081] cfDNA methylation feature extraction Cell-free DNA was extracted from the sample using the plasma sample extraction kit and the extraction steps as follows: 1. Transfer 4 ml of plasma sample to pre-labeled 15 ml centrifuge tubes; add 6 ml of lysis buffer DL. Tightly cap the centrifuge tubes and vortex to mix.

[0082] 2. Place the centrifuge tubes on a rotary mixer and rotate at medium speed (approximately 10-20 rpm) for 20 minutes at room temperature.

[0083] Note: Low-temperature storage of the kit may cause the lysis buffer DL to crystallize and precipitate. If crystallization of the lysis buffer DL is found before use, it can be dissolved in a 55°C water bath.

[0084] DNA binding 1. Add the following reagents to a 15ml centrifuge tube in sequence: 160µl magnetic beads (freshly suspended); tighten the cap of the centrifuge tube and invert to mix 5-6 times.

[0085] 2. Place the centrifuge tubes on a rotary mixer and rotate at medium speed (approximately 10-20 rpm) for 60 minutes at room temperature.

[0086] Thorough mixing of the magnetic beads is a key factor for successful experimentation; variations in the number of beads will lead to erroneous results. To ensure uniform concentration, the beads should be thoroughly mixed using a vortex mixer before pipetting until no obvious sediment remains at the bottom of the flask. Simultaneously, the beads must be kept in uniform suspension during pipetting.

[0087] DNA washing 1. Place a 15ml centrifuge tube on a magnetic rack for 1-3 minutes to allow the magnetic beads to adhere. Carefully discard the supernatant (do not discard the magnetic beads; keep the 15ml centrifuge tube on the magnetic rack while discarding the supernatant). If the magnetic beads in the plasma sample cannot be effectively separated by the magnetic rack, remove the 15ml centrifuge tube from the above steps, add 2.5ml of lysis buffer, invert and mix 5-6 times, and place it back on the magnetic rack for further separation. If the magnetic beads still cannot be effectively separated, place the 15ml centrifuge tube in a 55℃ (50-60℃) water bath for 10 minutes, then place it on the magnetic rack for another round of separation.

[0088] 2. Add 1.0 ml of washing solution A and vortex to ensure complete resuspension of the magnetic beads. Use a disposable pipette to transfer the magnetic bead suspension to a labeled 1.5 ml centrifuge tube. Use a pipette again to aspirate any remaining magnetic bead suspension and transfer it to a 1.5 ml centrifuge tube.

[0089] 3. Place the 1.5ml centrifuge tube on the magnetic rack for 1 minute to allow it to adhere. Carefully aspirate all the liquid, being careful not to aspirate the magnetic beads.

[0090] 4. Briefly centrifuge the 1.5ml centrifuge tube. Place the 1.5ml centrifuge tube on a magnetic rack to re-adhere for 1 minute, and use a 10-100µl pipette to remove as much residual liquid as possible.

[0091] 5. Let the centrifuge tubes air dry at room temperature for 10 minutes. Washout 1. Add 45µl of elution buffer. Cover the centrifuge tube, vortex to mix, and resuspend the magnetic beads.

[0092] 2. Place the centrifuge tubes in a constant temperature shaker and shake at 1000±100 rpm for 10 minutes at room temperature.

[0093] 3. Briefly centrifuge the centrifuge tubes by placing them on a magnetic rack for 1 minute.

[0094] 4. Use a 10-100µl pipette to transfer 42µl of elution solution into a new centrifuge tube.

[0095] Qubit concentration measurement Take 1 μL to measure the concentration and determine the starting amount.

[0096] DNA treatment with bisulfite Starting amount: 10-200 ng; a starting amount greater than 40 ng can increase region coverage; volume: 40 μL; DNA Protect Buffer changes from green to blue to indicate thorough mixing and correct pH. See Table 1: Table 1 The reaction conditions, Lid 105℃, are shown in Table 2: Table 2 In some cases, the incubation time at 60°C can be extended to 20 minutes to achieve complete conversion.

[0097] Column purification 1. Add anhydrous ethanol to Buffer BW, Buffer BW, and Buffer BD, and add ribonuclease-free water to the vector RNA.

[0098] 2. After conversion, centrifuge and transfer the product to a 1.5 ml centrifuge tube. The transfer of the precipitate will not affect the reaction.

[0099] 3. Add 310 μL of freshly prepared Buffer BL containing 10 μg / ml vector RNA, mix well, and centrifuge. When the initial amount is >40 ng, vector RNA is not required. The configuration of Buffer BL is shown in Table 3: Table 3 4. Add 250 μL of anhydrous ethanol, mix well for 15 seconds, and centrifuge.

[0100] 5. Prepare the appropriate number of columns and transfer all the reaction solution from the previous step into the columns.

[0101] 6. Centrifuge at 14000 rpm for 1 min and discard the effluent.

[0102] 7. Add 500 μL of Buffer BW to the column, centrifuge at 14,000 rpm for 1 min, and discard the eluent.

[0103] 8. Add 500 μL of Buffer BD (desulfonated) to the column, cap immediately to prevent acidification caused by carbon dioxide in the air, and incubate at room temperature for 15 minutes. Buffer BD easily forms a precipitate, so avoid transferring the precipitate into the column.

[0104] 9. Centrifuge at 14000 rpm for 1 min and discard the effluent.

[0105] 10. Add 500 μL of Buffer BW to the column, centrifuge at 14,000 rpm for 1 min, and discard the eluent. 11. Repeat the above step.

[0106] 12. Add 250 μL of anhydrous ethanol and centrifuge at 14,000 rpm for 1 min.

[0107] 13. Place the column into a new 2ml collection tube, centrifuge at 14000rpm for 1min, and discard the eluent.

[0108] Optional: Open the column and place it into a new 1.5ml centrifuge tube, then incubate in a metal bath at 60℃ for 5 minutes.

[0109] 14. Place the column into a 1.5ml centrifuge tube, add 22ul of Buffer EB to the center of the membrane, cap it, and incubate at room temperature for 1min.

[0110] 15. Centrifuge at 12000 rpm for 1 min.

[0111] Purified DNA can be stored at 2-8℃ for 24 hours. If the storage time exceeds 24 hours, it should be stored at -20℃. Sulfite conversion of less than 200ng starting amount should not exceed 2 weeks.

[0112] cfDNA methylation library construction: End repair is shown in Table 4: Table 4 The reaction conditions were as follows: placed on a PCR instrument, 37℃ for 30 min; 65℃ for 20 min; 4℃ hold (Lid 75℃).

[0113] The connector connections are shown in Table 5: Table 5 Reaction conditions: Place on a PCR instrument at 25°C for 15 min; maintain at 4°C (heating lid closed); The QIAseq Methylated DNA 8-index kit contains 8 individually packaged IL-Me-N7## adapters. The accompanying library amplification primers only provide IL-S502 index primers. Only one adapter needs to be opened for each experiment to avoid cross-contamination. The QIAseq Methylated DNA 96 Index Kit (Series A, B, C, or D) contains 12 IL-Me-N7## adapters per plate, 8 μl per well. Thaw and centrifuge before use. The accompanying library amplification primers are the corresponding IL-S5 index primer plates (Series A, B, C, or D), each plate providing 8 index primers. Both the IL universal primers and the primers are provided in dry powder form.

[0114] Ligation product purification Add 144 μl (1.6 times the volume) of QIAseq magnetic beads for purification (the beads need to equilibrate at room temperature for 30 minutes), mix thoroughly until no visible clumps remain, and then incubate at room temperature for 5 minutes. After a brief centrifugation, place the tube on a magnetic rack and let it stand for 5 minutes, then discard the supernatant. On the magnetic rack, wash the magnetic bead-nucleic acid complex with 200 μl of 70% ethanol, discard the ethanol, and then briefly centrifuge. Carefully aspirate any remaining liquid using a small volumetric pipette tip. After drying at room temperature for 5-10 minutes, remove the tube from the magnetic rack, add 55 μl of nuclease-free water, mix well, and incubate at room temperature for 5 minutes. After a brief centrifugation, place the tube back on the magnetic rack and let it stand for 5 minutes, then aspirate 52 μl of the supernatant into a new 1.5 ml centrifuge tube.

[0115] Add 83.2 μL (1.6×) QIAseq magnetic beads for purification, mix well, incubate at room temperature for 5 min, briefly centrifuge, place on a magnetic rack for 5 min, and discard the supernatant. Wash with 200 μL 70% ethanol (on a magnetic rack), discard the liquid, briefly centrifuge, and aspirate dry with a pipette tip. After drying at room temperature for 5-10 min, remove from the magnetic rack, add 20 μL nuclease-free water, mix well, incubate at room temperature for 5 min, briefly centrifuge, place on a magnetic rack for 5 min, and transfer 17 μL to a PCR tube. Can be stored overnight at -20°C.

[0116] The target region enrichment is shown in Table 6: Table 6 The reaction conditions, Lid 105℃, are shown in Table 7: Table 7 After completion, add 2 μL of TM stop solution to terminate the reaction. If necessary, store overnight at -20°C.

[0117] Product purification Add 67.2 μL (1.6×) QIAseq magnetic beads for purification, mix well, and incubate at room temperature for 5 min until no visible clumps form. Briefly centrifuge, place on a magnetic rack for 5 min, and discard the supernatant. Wash with 200 μL 70% ethanol (on a magnetic rack), discard the liquid, and briefly centrifuge. Aspirate dry with a pipette tip. Dry at room temperature for 5-10 min, remove from the magnetic rack, add 55 µL nuclease-free water, mix well, incubate at room temperature for 5 min, briefly centrifuge, place on a magnetic rack for 5 min, and transfer 52 µL of supernatant to a new tube.

[0118] Add 83.2 μL (1.6×) QIAseq magnetic beads for purification, mix well, incubate at room temperature for 5 min, briefly centrifuge, place on a magnetic rack for 5 min, and discard the supernatant. Wash with 200 μL 70% ethanol (on a magnetic rack), discard the liquid, briefly centrifuge, and aspirate dry with a pipette tip. After drying at room temperature for 5-10 min, remove from the magnetic rack, add 20 μL nuclease-free water, mix well, incubate at room temperature for 5 min, briefly centrifuge, place on a magnetic rack for 5 min, and transfer 17 μL to a PCR tube.

[0119] Library expansion is shown in Table 8: Table 8 If using the QIAseq methylated DNA 96-index kit (series A, B, C, or D), the library amplification step must be performed using the corresponding series of IL-S5 index primers (Set A, B, C, or D). Both the IL-S5 index primers and the universal IL primers are pre-aliphatic in the wells and must be centrifuged before use. Prepare 6.6 µl of universal PCR mixture, consisting of: 1.6 µl nuclease-free water, 4 µl QIAseq UPCR buffer (5x), and 1 µl hot-start Taq DNA polymerase. Then, add 13.4 µl of the purified DNA from the previous step to the corresponding well of the 96-well primer plate, followed by 6.6 µl of the prepared PCR mixture, bringing the total volume to 20 µl. Finally, transfer the reaction mixture from the wells to PCR tubes for subsequent library amplification.

[0120] Reaction conditions: Lid 105℃, see Table 9: Table 9 See Table 10 for the number of cycles: Table 10 Purification of library amplification products: Add 80 μL of nuclease-free water (placed on ice) to the product from the previous step, mix well, add 160 μL of (1.6×) QIAseq magnetic beads for purification, mix well, incubate at room temperature for 5 min, briefly centrifuge, place on a magnetic rack for 5 min, and discard the supernatant. Wash with 200 μL of 70% ethanol (on a magnetic rack), discard the liquid, briefly centrifuge, and aspirate dry with a pipette tip. After drying at room temperature for 5-10 min, remove from the magnetic rack, add 25 μL of nuclease-free water, mix well, incubate at room temperature for 5 min, briefly centrifuge, place on a magnetic rack for 5 min, and transfer 20 μL to a new tube.

[0121] Document Quality Control: For concentration measurement using Qubit, the concentration must be >2 ng / ul. The Agilent 2100 Bioanalyzer was used for analysis. The library was diluted 1:2 and loaded onto an Agilent high-sensitivity DNA chip for analysis. A qualified library should have fragments concentrated between 200-600 bp, with the main peak located at approximately 350 bp. This main peak consists of a core fragment of approximately 170 bp of circulating cell-free DNA plus a sequencing adapter of approximately 180 bp.

[0122] Second-generation sequencing The methylation sequencing of the amplified library products was performed using an Illumina sequencer to obtain sequencing results.

[0123] Off-machine data analysis process: UMI tag: The first 12 bases of Read2 are the molecular tag UMI sequence. FastP is called to treat each pair of FASTQ files as paired reads and add UMI tags and low-quality bases to generate FASTQ files with added UMI tags.

[0124] Quality control connector removal: Use cutadapt to remove the connector. If the read length is less than 50 after removing the connector, discard it.

[0125] Alignment and deduplication: BisMark is called to align each pair of fastq files as paired reads to the hg19 human reference genome sequence, generating the initial Bam file; Call deduplicate_bismark to remove duplicate sequences from the initial Bam file based on the marked UMI sequences; Sorting: Use Samtools to sort the initial Bam file (after removing duplicate sequences) according to chromosome position; Filtering: The view command in Samtools is used to filter the above Bam files, filtering out reads with low alignment quality (the alignment quality must be above 20) to generate the final Bam file. An internal script is used to filter reads where the non-CpG CT conversion rate is below 95% on each read. Create an index: Call the index module of Samtools to create an index for the final generated Bam file, and generate bai files that are paired with the Bam files after the tags are duplicated.

[0126] Calculate the average methylation level (AMF) for each methylation block (MB). Calculate the Beta value (methylation level at a single point): Use BisSNP to obtain the Beta value for each CpG site; The average methylation level (AMF) of CpG points contained in each MB is calculated and used as the average methylation level (AMF) of each MCB.

[0127] Clinical feature extraction Clinical data were collected from patients, including demographic information (age, sex, smoking history), tumor markers (carbohydrate antigen 125, squamous cell carcinoma-associated antigen, CEA, cytokeratin 19 fragment, etc.), and hematological parameters (lymphocyte count, platelet count, lactate dehydrogenase, neutrophil-to-lymphocyte ratio (NLR), etc.). Univariate analysis (P<0.05) and LASSO regression were used to screen for and retain six key clinical characteristics: age ≥60 years, smoking history ≥20 pack-years, CEA ≥5 ng / mL, squamous cell carcinoma-associated antigen ≥1.5 ng / mL, NLR ≥3, and platelet count ≥300 × 10⁻⁶. 9 / L.

[0128] CT image feature extraction Image acquisition and annotation: Nodule images (slice thickness 1.25mm) were acquired using a 128-slice CT scanner and saved in DICOM format; the nodule region was manually segmented by two radiologists with more than 5 years of experience using MITK software, with a Kappa coefficient ≥0.85 (in case of disagreement, a third physician would arbitrate).

[0129] Feature extraction and selection: 1800 radiomics features (covering dimensions such as morphology, texture, and grayscale) were extracted using the Pyradiomics tool. The intraclass correlation coefficient (ICC) was calculated, and 79 stable features with an ICC ≥ 0.8 were retained as CT image features.

[0130] Diagnostic model construction and validation Single-modal model construction The methylation model (Solid_Pulmonary_Methylation_Model) uses 120 methylation features as input and employs a logistic regression algorithm to construct the model. Parameter settings are: L2 penalty, 1e-4 tolerance, reciprocal of regularization strength C=0, True intercept, lbfgs solver, and a maximum of 100 iterations. The model is optimized using 5-fold cross-validation (Stratified KFold). The average AUC on the training set is 0.94±0.02, and the AUC on the validation set is 0.97 (95% CI: 0.94–1.0).

[0131] Clinical Model (Solid_Pulmonary_Clinical_Model): The model was constructed using a logistic regression algorithm with 6 clinical features as input, and optimized with 5-fold cross-validation. The validation set AUC was 0.76 (95% CI: 0.64-0.83).

[0132] CT Model (Solid_Pulmonary_CT_Model): The model is constructed using a logistic regression algorithm with 79 CT image features as input, and optimized with 5-fold cross-validation. The validation set AUC is 0.78 (95% CI: 0.65-0.91).

[0133] Multimodal model construction (Solid_Pulmonary_Multi_Model) Feature fusion: Z-score normalization was applied to methylation features, min-max normalization was applied to clinical features, and L2 regularization was applied to CT features; 130 key features (including methylation, clinical, and CT features) were selected by SHAP value, and low-rank matrix factorization (rank r=30) was used to unify feature dimensions.

[0134] Model training and validation: A fusion model was constructed using logistic regression, optimized with cross-entropy loss function, and tuned using 5-fold cross-validation. Independent validation set results showed an AUC of 0.97 (95% CI: 0.94–1.0), sensitivity of 83.81%, specificity of 100%, and accuracy of 88.98%, significantly outperforming the single-modal model.

[0135] The method for identifying solid lung nodules based on multimodal data fusion in this application combines methylation features, CT image features, and other features to construct a multimodal fusion model for training and validation. Through validation on the test set, the model can output the predicted probability of benign and malignant nodules, thereby greatly improving the prediction accuracy.

[0136] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0137] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0138] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0139] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for identifying solid lung nodules based on multimodal data fusion, characterized in that, The method includes: Acquire multimodal data of lung nodules; Construct the training and validation sets for the multimodal data; For the training set, the blood samples, CT image data and clinical data in the multimodal data are preprocessed respectively to obtain methylation features, CT image features and clinical features in sequence; A multimodal fusion model is constructed based on the methylation features, the CT image features, and the clinical features. The multimodal fusion model is validated based on the validation set. Once the verification is successful, the predicted probability of benign or malignant nodules is obtained through the multimodal fusion model.

2. The method according to claim 1, characterized in that, The construction of the training and validation sets for the multimodal data specifically includes: Acquire multimodal data from patients with solid nodules, pure ground-glass nodules, and mixed ground-glass nodules; A first number of multimodal data points and a second number of multimodal data points are obtained from the multimodal data points respectively; the number of solid nodules, pure ground-glass nodules, and mixed ground-glass nodules in the first number of multimodal data points conforms to a first ratio; the number of solid nodules, pure ground-glass nodules, and mixed ground-glass nodules in the second number of multimodal data points conforms to a second ratio; the first number is not equal to the second number; the first ratio is equal to or not equal to the second ratio; Based on the first multimodal data, construct a training set; Based on the second multimodal data, construct a validation set.

3. The method according to claim 1, characterized in that, The obtained methylation features specifically include: Based on the preset methylation panel, the blood sample to be tested is captured, predicted, and preprocessed to obtain a Bam file; the Bam file records the precise location of each sequencing read aligned to the reference genome, whether the site is methylated, and whether the methylation of the site changes synchronously. Based on preset rules and the Bam file, methylation blocks are divided; the number of cytosine-guanine dinucleotide (CpG) sites in the same methylation block is >2, and the Pearson correlation coefficient of any adjacent CpG sites is >0.5; Based on the Bam file, calculate the number of methylated reads and the total number of reads for each CpG site within each methylation block; The average methylation level (AMF) of each methylated block is obtained by summing the methylated reads of all sites and dividing by the total number of reads of all sites. Extract each individual sequencing sequence from the Bam file; Each sequencing sequence is scored; for a sequencing sequence with k CpG sites, the weight is the first value when all sites are unmethylated, the weight is the second value when all sites are methylated, and the weight is m / k when m consecutive sites are methylated. The methylation haplotype load (MHL) is calculated as MHL = Σ(weight of each sequence) / total number of sequences; where the total number of sequences refers to the total number of sequencing sequences covering the methylation block. Calculate the Pearson correlation coefficient between each AMF or MHL feature and the benign / malignant label, and select the initial methylation features based on the Pearson correlation coefficient. Methylation features were determined from the initial methylation features using the LASSO regression method.

4. The method according to claim 3, characterized in that, The process of obtaining the Bam file also includes: Blood samples were obtained from multimodal data of lung nodules and blood samples from controls; cfDNA was extracted, sequencing libraries were constructed, and enzymatic transformation was performed to obtain amplified library products; The amplified library products were sequenced using a high-throughput sequencer to output the raw FASTQ file; the FASTQ file includes multiple DNA sequences and their quality information. The original FASTQ file is processed, and fastp is called to add UMI tags to paired reads, filtering out bases that do not meet the quality requirements, and generating a FASTQ file with UMI tags; wherein, filtering out bases that do not meet the quality requirements includes removing the adapter with the cutadapt tool and discarding reads with a length <50bp; the first 12 bases of Read2 are UMI tags; The BisMark tool was used to align the filtered paired reads to the hg19 human reference genome to generate the initial Bam file. Based on the UMI tag, repetitive sequences introduced by PCR amplification were removed. The initial Bam file was sorted by chromosome position using Samtools and reads with an alignment quality >20 were selected. Reads with a CT conversion rate <95% at non-CpG sites were filtered to generate the Bam file and create an index.

5. The method according to claim 1, characterized in that, The obtained CT image features specifically include: Obtain the first region of interest (ROI) and the second region of interest (ROI) from the CT image data determined by the first terminal and the second terminal, respectively. The consistency between the first ROI and the second ROI was verified using the Kappa coefficient; Once the verification is successful, the target ROI is determined. Extract the radiomics features of the target ROI; Calculate the intraclass correlation coefficient (ICC) value of the radiomics features; The initial radiomics features corresponding to ICC values ​​greater than the first threshold are retained; The variance of each initial radiomics feature is calculated, and the variance is compared with a preset second threshold. The initial CT image features with variances greater than the second threshold are retained as CT image features.

6. The method according to claim 5, characterized in that, The CT image features include morphological features, grayscale features, texture features, and higher-order features; the morphological features describe the geometry of the nodules, the grayscale features describe the distribution of pixel values ​​within the ROI, the texture features describe the spatial distribution and relationship of pixel grayscale values, the relationship including contrast and correlation, and the higher-order features are features extracted after filtering the image.

7. The method according to claim 1, characterized in that, The obtained clinical characteristics specifically include: Clinical data are constructed based on demographic information, medical history information, tumor markers, hematological parameters, and derived indicators. Each variable in the clinical data is processed to determine the difference of each variable between the benign and malignant groups; the processing includes using the T test or Mann-Whitney U test for continuous variables, the chi-square test for categorical variables, and the rank-sum test for ordinal variables; Variables with p-values ​​less than the third threshold are retained to obtain the first variable; For the first variable, the second variable is obtained by screening using LASSO regression and Logistic regression with minimum absolute contraction and selection operators. The second variable is binarized according to the critical value to obtain a binary classification feature, which is a clinical feature.

8. The method according to claim 1, characterized in that, The construction of the multimodal fusion model based on the methylation features, the CT image features, and the clinical features specifically includes: Z-score normalization was applied to methylation features, normalization was applied to clinical features, and L2 regularization was applied to CT image features. The SHAP values ​​of processed methylation features, CT imaging features, and clinical features were evaluated, and the features were sorted according to their SHAP values, with the top n features selected. The first n features are decomposed using a low-rank matrix to obtain the decomposition results; Element-wise product fusion is performed using F=∏(M_i×W_i) to fuse all mode decomposition results; where ∏ is the element-wise product operation, M_i is the feature matrix of the i-th mode, and W_i is the mode weight matrix; Logistic regression is used as the base model, support vector machine is used as the meta-model, and they are stacked and integrated. The multimodal fusion model is then optimized using the cross-entropy loss function to obtain the multimodal fusion model.

9. The method according to claim 8, characterized in that, The method further includes: The overall performance of the multimodal fusion model was verified using an independent test set; wherein the overall performance was measured by the area under the ROC curve (AUC), sensitivity, specificity, and accuracy.

10. The method according to claim 1, characterized in that, The method further includes: Based on the methylation features and validation set, the CT image features and validation set, and the clinical features and validation set, respectively, a methylation feature model, a CT image feature model, and a clinical feature model are constructed: The training process involves using multiple methylation features of each sample in the training set and the labels corresponding to those methylation features as input to train a methylation feature model. The performance of the methylation feature model is evaluated using 5-fold cross-validation, and the parameters are adjusted until optimal. The methylation feature model is then evaluated using a test set, and the training ends when AUC ≥ 0.

9. The specific steps of constructing a CT image feature model based on CT image features are as follows: taking multiple CT image features of each sample in the training set and the labels corresponding to the CT image features as input, training the CT image feature model, evaluating the performance of the CT image feature model using 5-fold cross-validation, adjusting the parameters until optimal, evaluating the CT image feature model through the test set, and ensuring that AUC ≥ 0.9, at which point the training of the CT image feature model is complete. The specific steps of constructing a clinical feature model based on clinical characteristics include: taking multiple clinical features of each sample in the training set and the labels corresponding to the clinical features as input, training the clinical feature model, evaluating the performance of the clinical feature model using 5-fold cross-validation, adjusting the parameters until optimal, evaluating the clinical feature model through the test set, and completing the training of the clinical feature model when AUC ≥ 0.

9. Based on the outputs of the methylation feature model, CT image feature model, and clinical feature model, a multimodal fusion model is constructed.

Images