A protein-medical record-image multimodal lung nodule screening system and storage medium
By constructing a multimodal lung nodule screening system that combines imaging and protein information, the problems of high false positive rate and high cost in lung nodule screening have been solved, achieving efficient and accurate lung nodule screening.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XINJIANG PROD & CONSTR CORPS HOSPITAL (SECOND AFFILIATED HOSPITAL OF SHIHEZI UNIV MEDICAL COLLEGE)
- Filing Date
- 2026-03-31
- Publication Date
- 2026-06-09
AI Technical Summary
Existing technologies for screening lung nodules suffer from weak image-protein correlation, baseline protein shift in the population, and a trade-off between screening costs and efficiency. Multimodal information fusion models also exhibit poor robustness, leading to high false positive rates, high missed diagnoses, and high costs.
A multimodal fusion strategy of protein-medical record-image is adopted. By constructing a local population multimodal reference database and combining image analysis, protein correction and risk calculation, data-driven multi-level linkage screening is realized. This includes data collection, multimodal reference database, intelligent decision engine and feedback optimization module to optimize image score weight and disease-specific threshold.
It improves the accuracy of screening results, reduces false positive rates and screening costs, ensures the consistency and interpretability of results, and is suitable for large-scale population screening.
Smart Images
Figure CN122176314A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical software, specifically relating to a protein-medical record-image multimodal lung nodule screening system and its storage medium. Background Technology
[0002] Early qualitative diagnosis of pulmonary nodules is a crucial step in the prevention and treatment of lung cancer. The widespread adoption of low-dose CT screening has significantly increased the detection rate of pulmonary nodules; however, many indeterminate nodules still rely on follow-up observation or invasive biopsy. The former may delay diagnosis, while the latter carries invasive risks. Single-modal diagnostic information has limitations: radiomics can provide morphological features, but suffers from the problem of "different images for the same disease, and the same image for different diseases"; protein biomarkers can reflect tumor biological behavior, but are greatly affected by background interference; clinical variables provide background risk, but their sensitivity is insufficient when used alone. Integrating multimodal information holds promise for achieving complementary advantages. However, directly applying this model to lung nodule screening faces the following prominent problems: 1. Weak image-protein correlation: Relying solely on image screening has a high false positive rate, leading to many patients with benign nodules undergoing unnecessary follow-up or invasive examinations, while relying solely on protein testing is prone to missing some early lung cancers with atypical imaging features; 2. Population protein baseline drift: There are significant differences in the baseline levels of protein markers among people of different ages, sexes, and smoking histories, and directly applying a uniform reference value can lead to an increased false positive rate; 3. Contradiction between screening cost and efficiency: Comprehensive omics testing (such as whole genome sequencing and multi-omics testing) is costly and difficult to promote in large-scale populations at this stage; while relying solely on image screening is prone to missing diagnoses.
[0003] While some existing technologies have attempted to integrate multimodal information, they mostly employ simple feature splicing, failing to adequately consider the alignment issues of heterogeneous data between modalities and lacking dynamic adjustment mechanisms for feature contributions. This results in poor robustness and weak interpretability of the fusion model. Therefore, there is an urgent need for a data-driven, multi-level, and continuously self-optimizing multimodal screening decision system specifically designed for lung nodule screening. Summary of the Invention
[0004] To address the problems existing in the prior art, the present invention provides a protein-medical record-image multimodal lung nodule screening system and storage medium.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a protein-medical record-image multimodal lung nodule screening system, comprising a data acquisition module, a multimodal reference database, and an intelligent decision engine.
[0006] The data acquisition module is used to collect low-dose CT images of the lungs, plasma protein biomarker test values, and clinical medical record information of the examinee and transmit the information to the multimodal reference database.
[0007] The multimodal reference database pre-stores reference ranges for radiomics features, reference medians for protein biomarkers, and reference standard deviations based on the local pulmonary nodule population.
[0008] The intelligent decision engine includes an image analysis unit, a protein correction unit, and a risk calculation unit. The image analysis unit is used to detect and segment nodules in low-dose CT images, extract radiomics features, and output a preliminary risk score S_img for benign or malignant nodules. The protein correction unit is used to retrieve the reference median and reference standard deviation of the corresponding protein biomarkers from a multimodal reference database based on the patient's clinical medical record information (age, gender, smoking history) when the image analysis unit outputs a positive nodule, and calculate the corrected Z-score of the protein biomarker detection value. The risk calculation unit calculates the comprehensive risk index CRI based on the preliminary risk score S_img, the corrected Z-score, and the preset disease-specific normalized weights.
[0009] Preferably, the data acquisition module also records the subject's unique identifier, examination institution, examination time, age, gender, smoking history, nodule size, and nodule density. The system automatically parses the CT images through the DICOM interface and binds the above information with the test results to the cloud.
[0010] Preferably, the multimodal reference database also pre-stores a population-specific mutation database, which pre-stores pathogenic or potentially pathogenic mutation sites of lung nodule-related genes.
[0011] Preferably, the image analysis unit uses a 3D convolutional neural network to perform end-to-end feature extraction on CT images and outputs a benign or malignant probability score. If the image analysis unit does not detect nodules, the system directly outputs a negative report without triggering subsequent detection.
[0012] Preferably, the intelligent decision engine further includes a report generation module and a feedback optimization module. The report generation module generates a structured lung nodule screening risk assessment report according to the relationship between the comprehensive risk index and the preset disease-specific threshold, and according to the preset positive / negative judgment rules. The feedback optimization module receives clinical diagnosis result data, periodically refits the reference range of radiomics features and the reference median and reference standard deviation of protein biomarkers, optimizes the image score weight, disease-specific weight and disease-specific threshold, and updates the multimodal reference database and the population-specific mutation database.
[0013] Preferably, the construction steps of the multimodal reference database are as follows:
[0014] S1: Reference range of radiomic features: Collect CT images of patients with pathologically confirmed pulmonary nodules, extract radiomic features (including first-order statistics, shape, texture, and wavelet features), stratify them according to nodule size (≤10mm, 10-20mm, >20mm) and nodule density (solid, partially solid, ground-glass), statistically analyze the distribution of imaging features of benign and malignant nodules respectively, and establish the reference range of imaging features of benign and malignant nodules in each subgroup;
[0015] S2 protein biomarker reference median and reference standard deviation: Based on stratification by age (≤50 years, >50 years), sex (male, female), and smoking history (yes, no), at least 50 healthy controls and benign nodules were recruited in each subgroup. Plasma samples were collected, and the concentrations of target protein biomarkers (PCSK9, PDIA3, MUC1 fragment, CEA, CYFRA21-1) were detected. The median and standard deviation were calculated for each subgroup. Continuous feature-reference median curves and feature-reference standard deviation curves were established using local weighted regression (LOESS). The fitting parameters were set as follows: span=0.5, degree=1.
[0016] Preferably, the mutation sites in the population-specific mutation database include one or more of the following sites: p.L858R and exon 19 deletion of the EGFR gene; p.R273H and p.R175H of the TP53 gene; p.G12C and p.G12D of the KRAS gene; and phosphorylation-related mutations at the S408 site of the FGFR3 gene.
[0017] Preferably, the disease-specific weights and disease-specific thresholds are determined as follows: Imaging screening risk scores, protein biomarker detection values, and clinical medical record information of previously confirmed cases and healthy controls are collected. The diagnosis result (malignant = 1, benign = 0) is used as the dependent variable, and the imaging screening risk score and the corrected Z-scores of each protein biomarker are used as independent variables. A logistic regression algorithm is used for training the model. 10-fold cross-validation is used during model training to prevent overfitting. The regression coefficients obtained after training are normalized and used as disease-specific weights. The optimal threshold is determined based on maximizing the Youden exponent, and this threshold is used as the disease-specific threshold.
[0018] The corrected Z-score for protein biomarker detection values is calculated using the following formula:
[0019]
[0020] Where X is the measured protein biomarker value, M(h) is the reference median of the protein biomarker in the corresponding clinical characteristic subgroup, and SD(h) is the reference standard deviation of the protein biomarker in the corresponding clinical characteristic subgroup.
[0021] The overall risk index is calculated using the following formula:
[0022]
[0023] Wherein, CRI is the comprehensive risk index, S_img is the initial image screening risk score, w_img is the image score weight, Z_i is the corrected Z-score of the i-th protein biomarker, w_i is the disease-specific normalized weight of the i-th protein biomarker, and s_i is the risk direction sign of the i-th protein biomarker, which takes the value of +1 (indicating that the larger the Z-score, the higher the risk) or -1 (indicating that the smaller the Z-score, the higher the risk). When s_i = -1, the negative value of Z_i is mapped to a positive value to participate in risk accumulation, and all weights satisfy the normalization condition.
[0024] The present invention also provides a computer-readable storage medium, wherein the computer program, when executed by a processor, runs the system as described in any one of claims 1-8.
[0025] This invention provides a protein-medical record-image multimodal lung nodule screening system and storage medium. It has the following beneficial effects:
[0026] This invention abandons the single-modal screening model and adopts a three-modal fusion strategy of imaging-protein-clinical. By constructing a local population multimodal reference database, it can improve the accuracy of screening results and reduce the false positive rate.
[0027] Built-in protein biomarker reference curves based on age, gender, and smoking history address the protein baseline drift problem and improve screening specificity.
[0028] By implementing a three-tiered control system of initial imaging screening, protein confirmation, and risk scoring, protein testing is initiated only when nodules are detected by imaging, which can reduce the proportion of protein testing and significantly reduce the overall screening cost.
[0029] The feedback optimization module enables the system's reference thresholds and weights to continuously iterate as confirmed data accumulates, eliminating the need for frequent manual adjustments.
[0030] This invention specifies in detail the content structure of the lung nodule screening risk assessment report and the rules for determining positive / negative results, ensuring the consistency and interpretability of the system output.
[0031] In summary, this invention constructs a complete multimodal screening system and process for pulmonary nodules, optimized for the clinical characteristics of the screening population, thus improving screening accuracy. Simultaneously, it introduces an initial imaging screening step, which effectively reduces the proportion of protein detection, thereby better meeting the objective needs of large-scale screening. Therefore, this invention combines multimodal fusion, precise screening, and cost control, demonstrating promising clinical application prospects. Attached Figure Description
[0032] Figure 1 This is a system architecture diagram of the present invention. Detailed Implementation
[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0034] Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.
[0035] Please see Figure 1 This invention provides a technical solution: a protein-medical record-image multimodal lung nodule screening system, including a data acquisition module for acquiring low-dose CT images of the lungs, plasma protein biomarker test values, and clinical medical record information of the examinee. The data acquisition module also records the examinee's unique identifier, examination institution, examination time, age, gender, smoking history, nodule size, and nodule density. The system automatically parses the CT images via a DICOM interface and binds the above information with the test results, uploading them to the cloud.
[0036] A multimodal reference database is pre-stored with reference ranges for radiomics features, reference medians and reference standard deviations for protein biomarkers, constructed based on the local pulmonary nodule population.
[0037] A population-specific mutation database pre-stores information on pathogenic or potentially pathogenic mutation sites in genes related to lung nodules.
[0038] The intelligent decision engine includes: an image analysis unit, a protein correction unit, and a risk calculation unit, wherein:
[0039] The image analysis unit is used for nodule detection and segmentation of low-dose CT images, extracting radiomics features, and outputting a preliminary risk score for benign or malignant nodules, S_img (0-100 points). The image analysis unit uses a 3D convolutional neural network to perform end-to-end feature extraction on CT images and outputs a benign / malignant probability score. If the image analysis unit does not detect a nodule, the system directly outputs a negative report without triggering subsequent detection.
[0040] The protein correction unit is used to retrieve the reference median and reference standard deviation of the corresponding protein marker from the multimodal reference database based on the patient's clinical medical record information (age, gender, smoking history) when the image analysis unit outputs a positive nodule, and calculate the corrected Z-score of the protein marker detection value.
[0041] The risk calculation unit calculates the comprehensive risk index (CRI) based on the initial image screening risk score S_img, the corrected Z-score, and the preset disease-specific normalization weights.
[0042] The report generation module generates a structured lung nodule screening risk assessment report based on the relationship between the comprehensive risk index and the preset disease-specific threshold, according to the preset positive / negative determination rules.
[0043] The feedback optimization module receives clinical diagnosis results data, periodically refits the reference range of radiomics features and the reference median and reference standard deviation of protein biomarkers, optimizes the image score weight, disease-specific weight and disease-specific threshold, and updates the multimodal reference database and population-specific mutation database.
[0044] The multimodal reference database is constructed as follows:
[0045] S1 Radiomics Feature Reference Range: CT images of patients with pathologically confirmed pulmonary nodules (300 malignant cases and 200 benign cases) were collected, and radiomics features (including first-order statistics, shape, texture, and wavelet features) were extracted. The nodules were stratified according to nodule size (≤10mm, 10-20mm, >20mm) and nodule density (solid, partially solid, ground-glass opacity). The distribution of imaging features of benign and malignant nodules was statistically analyzed separately, and the imaging feature reference range of benign and malignant nodules in each subgroup was established.
[0046] S2 protein biomarker reference median and reference standard deviation: Based on stratification by age (≤50 years, >50 years), sex (male, female), and smoking history (yes, no), at least 50 healthy controls and individuals with benign nodules were recruited in each subgroup. Plasma samples were collected, and the concentrations of target protein biomarkers (PCSK9, PDIA3, MUC1 fragment, CEA, CYFRA21-1) were measured. The median and standard deviation were calculated for each subgroup, and continuous feature-reference median and feature-reference standard deviation curves were constructed using locally weighted regression (LOESS). The fitting parameters were set as follows: span=0.5, degree=1.
[0047] In the population-specific mutation database, the mutation sites include one or more of the following sites: p.L858R and exon 19 deletion of the EGFR gene; p.R273H and p.R175H of the TP53 gene; p.G12C and p.G12D of the KRAS gene; and phosphorylation-related mutations at the S408 site of the FGFR3 gene.
[0048] The disease-specific weights and disease-specific thresholds were determined as follows: Imaging screening risk scores, protein biomarker test values, and clinical medical record information were collected from previously confirmed cases and healthy controls. The diagnosis result (malignant = 1, benign = 0) was used as the dependent variable, and the imaging screening risk score and the corrected Z-scores of each protein biomarker were used as independent variables. A logistic regression algorithm was employed for model training. 10-fold cross-validation was used during model training to prevent overfitting. The regression coefficients obtained after training were normalized and used as the disease-specific weights. The optimal threshold was determined based on maximizing the Youden exponent, and this threshold was used as the disease-specific threshold.
[0049] The corrected Z-score for protein biomarker detection values is calculated using the following formula:
[0050]
[0051] Where X is the measured protein biomarker value, M(h) is the reference median of the protein biomarker in the corresponding clinical characteristic subgroup, and SD(h) is the reference standard deviation of the protein biomarker in the corresponding clinical characteristic subgroup.
[0052] The overall risk index is calculated using the following formula:
[0053]
[0054] Wherein, CRI is the comprehensive risk index, S_img is the initial image screening risk score, w_img is the image score weight, Z_i is the corrected Z-score of the i-th protein biomarker, w_i is the disease-specific normalized weight of the i-th protein biomarker, and s_i is the risk direction sign of the i-th protein biomarker, taking a value of +1 (indicating that the larger the Z-score, the higher the risk) or -1 (indicating that the smaller the Z-score, the higher the risk). When s_i = -1, the negative value of Z_i is mapped to a positive value to participate in risk accumulation. All weights satisfy the normalization condition.
[0055] The screening risk assessment report includes the following:
[0056] Basic information of the examinee: including sample number, examination date, report generation date, age, gender, smoking history, nodule size, and nodule density;
[0057] Image analysis results include nodule location, size, density, and initial image screening risk score;
[0058] Protein biomarker detection results (e.g., triggers): include protein name, detection value, corrected reference range, Z-score, and anomaly marker;
[0059] Comprehensive risk index: including CRI value, disease-specific threshold, and risk level (low / medium / high).
[0060] Final conclusions include positive / negative results and follow-up recommendations.
[0061] The rules for determining positive / negative results are as follows:
[0062] situation Relationship between CRI and threshold Final conclusion Judgment basis 1 CRI ≥ threshold Positive High risk of multimodal integration 2 CRI < threshold Negative Low risk of multimodal integration 3 No nodules were detected on imaging. Negative No abnormalities in the imaging
[0063] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A protein-medical record-image multimodal pulmonary nodule screening system, characterized in that, It includes a data acquisition module, a multimodal reference database, and an intelligent decision engine; The data acquisition module is used to collect low-dose CT images of the lungs, plasma protein marker detection values, and clinical medical record information of the examinee and transmit the information to the multimodal reference database; The multimodal reference database pre-stores reference ranges for radiomics features, reference medians and reference standard deviations for protein biomarkers based on the local pulmonary nodule population. The intelligent decision engine includes an image analysis unit, a protein correction unit, and a risk calculation unit. The image analysis unit is used to detect and segment nodules in low-dose CT images, extract radiomics features, and output a preliminary risk score S_img for benign or malignant nodules. The protein correction unit is used to retrieve the reference median and reference standard deviation of the corresponding protein biomarkers from a multimodal reference database based on the patient's clinical medical record information (age, gender, smoking history) when the image analysis unit outputs a positive nodule, and calculate the corrected Z-score of the protein biomarker detection value. The risk calculation unit calculates the comprehensive risk index CRI based on the preliminary risk score S_img, the corrected Z-score, and the preset disease-specific normalized weights.
2. The protein-medical record-image multimodal lung nodule screening system according to claim 1, characterized in that: The data acquisition module also records the subject's unique identifier, examination institution, examination time, age, gender, smoking history, nodule size, and nodule density. The system automatically parses CT images through the DICOM interface and binds the above information with the test results to the cloud.
3. The protein-medical record-image multimodal lung nodule screening system according to claim 1, characterized in that: The multimodal reference database also contains a population-specific mutation database, which contains information on pathogenic or potentially pathogenic mutation sites of genes related to pulmonary nodules.
4. The protein-medical record-image multimodal lung nodule screening system according to claim 3, characterized in that: The image analysis unit uses a 3D convolutional neural network to perform end-to-end feature extraction on CT images and outputs a benign or malignant probability score. If the image analysis unit does not detect nodules, the system directly outputs a negative report and does not trigger subsequent detection.
5. The protein-medical record-image multimodal lung nodule screening system according to claim 4, characterized in that: The intelligent decision engine also includes a report generation module and a feedback optimization module. The report generation module generates a structured lung nodule screening risk assessment report according to the relationship between the comprehensive risk index and the preset disease-specific threshold and the preset positive / negative judgment rules. The feedback optimization module receives clinical diagnosis data, periodically refits the reference range of radiomics features and the reference median and reference standard deviation of protein biomarkers, optimizes the image score weight, disease-specific weight and disease-specific threshold, and updates the multimodal reference database and the population-specific mutation database.
6. The protein-medical record-image multimodal pulmonary nodule screening system according to claim 5, characterized in that, The steps for constructing the multimodal reference database are as follows: S1: Reference range of radiomics features: Collect CT images of patients with pathologically confirmed pulmonary nodules, extract radiomics features including first-order statistics, shape, texture, and wavelet features, and classify nodules by size ≤10mm, 10-20mm, >20mm and density (solid, partially solid, ground-glass) to statistically analyze the distribution of imaging features of benign and malignant nodules and establish reference ranges of imaging features for each subgroup of benign and malignant nodules; S2 protein biomarker reference median and reference standard deviation: Based on stratification by age (≤50 years, >50 years), sex (male, female), and smoking history (yes, no), at least 50 healthy controls and benign nodules were recruited in each subgroup. Plasma samples were collected, and the concentrations of target protein biomarkers (PCSK9, PDIA3, MUC1 fragment, CEA, CYFRA21-1) were detected. The median and standard deviation were calculated for each subgroup. Continuous feature-reference median curves and feature-reference standard deviation curves were established using local weighted regression (LOESS). The fitting parameters were set as follows: span=0.5, degree=1.
7. The protein-medical record-image multimodal lung nodule screening system according to claim 6, characterized in that: The population-specific mutation database contains one or more of the following sites: p.L858R and exon 19 deletion of the EGFR gene; p.R273H and p.R175H of the TP53 gene; p.G12C and p.G12D of the KRAS gene; and phosphorylation-related mutations at the S408 site of the FGFR3 gene.
8. The protein-medical record-image multimodal lung nodule screening system according to claim 7, characterized in that: The disease-specific weights and disease-specific thresholds are determined as follows: Imaging screening risk scores, protein biomarker test values, and clinical medical record information of previously confirmed cases and healthy controls are collected. The diagnosis result (malignant = 1, benign = 0) is used as the dependent variable, and the imaging screening risk score and the corrected Z-scores of each protein biomarker are used as independent variables. A logistic regression algorithm is used for training the model. 10-fold cross-validation is used during model training to prevent overfitting. The regression coefficients obtained after training are normalized and used as disease-specific weights. The optimal threshold is determined based on maximizing the Youden exponent, and this threshold is used as the disease-specific threshold. The corrected Z-score for protein biomarker detection values is calculated using the following formula: Where X is the measured protein biomarker value, M(h) is the reference median of the protein biomarker in the corresponding clinical characteristic subgroup, and SD(h) is the reference standard deviation of the protein biomarker in the corresponding clinical characteristic subgroup. The overall risk index is calculated using the following formula: Wherein, CRI is the comprehensive risk index, S_img is the initial image screening risk score, w_img is the image score weight, Z_i is the corrected Z-score of the i-th protein biomarker, w_i is the disease-specific normalized weight of the i-th protein biomarker, and s_i is the risk direction sign of the i-th protein biomarker, which takes the value of +1 (indicating that the larger the Z-score, the higher the risk) or -1 (indicating that the smaller the Z-score, the higher the risk). When s_i = -1, the negative value of Z_i is mapped to a positive value to participate in risk accumulation, and all weights satisfy the normalization condition.
9. A computer-readable storage medium, characterized in that; When the computer program is executed by the processor, it runs the system as described in any one of claims 1-8.