Thyroid nodule benign and malignant classification method based on clinical information, radiomics and gene detection

Abstract

The invention relates to the technical field of computer medical detection, in particular to a thyroid nodule benign and malignant classification method based on clinical information, radiomics and gene detection. According to the method, pathological diagnosis of thyroid nodules is used as a dependent variable, detection data including radiomics characteristics and molecular omics are used as continuous variables, and the detection data of at least three genes CLDN10, HMGA2 and LANM3 are selected as the data of the molecular omics; meanwhile, in combination with part of clinical pathological characteristics including gender, age and BRAF V600E mutation condition factors, a thyroid nodule preoperative diagnosis prediction model based on radiomics and molecular omics is constructed through an SVM modeling method. Gene detection, radiomics and clinical basic information are combined, and a thyroid nodule diagnosis model based on the combination is established on the basis of Chinese population. The method is used for assisting clinicians in distinguishing benign and malignant thyroid nodules and guiding clinical decisions

Description

Technical Field

[0001] The present invention relates to the technical field of computer-aided medical detection, and particularly to a method for classifying the benign and malignant of thyroid nodules based on clinical information, radiomics, and gene detection. Background Art

[0002] Thyroid cancer is the endocrine system malignant tumor with the fastest growing incidence rate in recent years. Papillary thyroid carcinoma (PTC) is the most common pathological type, accounting for more than 90% of all thyroid cancers. Accurately differentiating the benign and malignant of thyroid nodules before surgery is crucial. Accurate preoperative diagnosis can effectively assist thyroid cancer patients to obtain timely and effective treatment, and at the same time help patients with benign nodules avoid unnecessary surgeries, thereby avoiding corresponding surgical complications such as recurrent laryngeal nerve injury and hypoparathyroidism.

[0003] Fine needle aspiration biopsy (FNA) is the most commonly recommended method for preoperative diagnosis of thyroid nodules. The accuracy rate of FNA is about 72.8% to 87.2%, which is mainly related to histopathology. However, its limitations include the need for an experienced cytopathologist to make accurate judgments, and often uncertain cytological diagnosis results. It is reported that about 17% of FNA results are uncertain and cannot accurately distinguish between benign and malignant. Among these patients with uncertain FNA results, only about 5%-10% are ultimately malignant. Patients with uncertain cytological results often need to undergo repeated FNA or invasive diagnostic surgeries, which increases the potential risks of related postoperative complications and unnecessary surgical interventions.

[0004] In recent years, molecular detection has become a promising diagnostic tool for differentiating benign and malignant thyroid nodules. For example, a microarray-based test that combines the expression levels of 167 genes (Afirma® Gene Expression Classifier) can diagnose thyroid nodules with a sensitivity of 92% and a specificity of 52%. Another classifier called ThyroSeq v3 is a next-generation sequencing analysis based on DNA and RNA. By using next-generation sequencing, a sensitivity of 98%, a specificity of 81.8%, and a diagnostic accuracy rate of 90.9% can be achieved. Other studies, such as RosettaGX and ThyGenX / ThyraMIR, also show certain application prospects.

[0005] Although these methods provide certain help for nodule diagnosis, there are still limitations. First, the high cost limits their clinical application, especially in underdeveloped countries. Second, the current research objects are basically limited to European and American patients, and there are few similar studies on Asian patients. Due to the genetic background differences between different populations, the extrapolation performance of these results decreases.

[0006] Previous studies have shown that the ultrasound features of thyroid tumors, including tumor size, capsular invasion, and microcalcifications, are independent predictors of lymph node metastasis in PTC patients and have been widely used for screening thyroid nodules. However, this requires highly skilled sonographers, and visual inspection also has limitations and is somewhat subjective. Radiomics is an emerging technology that converts medical images into mineable data by extracting a large number of quantitative features from digital images. Radiomics embodies the pursuit of precision medicine, emphasizing providing the right treatment to the right patient at the right time. Currently, radiomics analysis has been applied to various diseases.

[0007] For example, the Chinese Patent Application for Invention (Publication No.: CN113436150A, Publication Date: September 24, 2021) discloses a method for constructing an ultrasound radiomics model for predicting the risk of lymph node metastasis, including: 1) Dataset construction: Obtain the original ultrasound medical images and clinical information of PTC patients, perform standardization and de-identification processing on the original ultrasound medical images and clinical information, and randomly divide the processed data into a training set and a validation set; 2) Region of interest (ROI) delineation: Delineate the region of interest in each ultrasound medical image; 3) Radiomics feature extraction: Extract high-dimensional radiomics features for each region of interest; 4) Radiomics analysis: Construct two radiomics labels for the training set respectively, and then combine the radiomics labels with the clinical information to construct a lymph node metastasis risk prediction model; 5) Prediction model validation.

[0008] For example, the Chinese Patent Application for Invention (Publication No.: CN116596890A, Publication Date: August 15, 2023) discloses a method for predicting the risk stratification of thyroid cancer based on dynamic images using graph convolutional networks, including the following steps; Step 1: Use image filtering enhancement and super-resolution networks to preprocess the ultrasound images, design an algorithm based on a graph convolutional neural network structure for thyroid ultrasound image segmentation, and establish a corresponding network model for training; Step 2: Use a feature extraction method of dynamic ultrasound radiomics based on artificial intelligence technology to extract features from the dynamic video sequence, and introduce deep learning methods to fuse the dynamic ultrasound radiomics to make the extracted radiomics features more abundant; Step 3: Use a multi-branch neural network model to integrate various different modality data, calculate multi-modal biomarkers by mining the complementary information of different modality data, and thus establish a risk stratification prediction model for thyroid cancer that can guide clinical decision-making; which can improve the accuracy and precision of thyroid cancer risk stratification.

[0009] As disclosed in a Chinese invention patent application (Publication No.: CN113223716A, Publication Date: August 6, 2021), a method for predicting the benign and malignant nature of cervical lymph nodes before ablation of papillary thyroid microcarcinoma includes the following steps: S1. Obtain gray-scale ultrasound images of cervical lymph nodes before minimally invasive ablation of papillary thyroid microcarcinoma; S2. Outline and segment the lymph node regions in the images; S3. Perform image matching and high-throughput feature extraction, and calculate omics features; S4. Perform feature dimensionality reduction; S5. Establish an imaging genomics prediction model and conduct external validation of the model; S6. Input the ultrasound images into the imaging genomics prediction model for prediction and output the prediction results, which can meet the clinical diagnosis and treatment needs, establish a prediction model for accurate prediction, and assist in accurate clinical decision-making.

[0010] As disclosed in a Chinese invention patent application (Publication No.: CN116746962A, Publication Date: September 15, 2023), a method for predicting lymph node metastasis in thyroid cancer includes obtaining ultrasonic imaging images of a patient's thyroid and thyroid clinical risk factors, extracting features from the ultrasonic imaging images of the thyroid to obtain ultrasonic features of the patient's thyroid, and inputting the ultrasonic features of the patient's thyroid and the thyroid clinical risk factors into a prediction model for lymph node metastasis in thyroid cancer to obtain a prediction result for lymph node metastasis in thyroid cancer. The present invention constructs a prediction model for lymph node metastasis in thyroid cancer through thyroid ultrasound imaging genomics features and clinical risk factors that may predict lymph node metastasis to predict lymph node metastasis.

[0011] However, in practice, the applicant divided 117 patients into a training group and a validation group. The data of the training set were used to screen features and construct an imaging genomics model, and the validation set was used for validation. Multiple feature selections were performed to identify reliable features and reduce redundancy. Based on the imaging features selected from the training set, six algorithms were used to construct an imaging genomics model for prediction: KNN, LR, DT, linear support vector machine, Gaussian SVM, and polynomial SVM. Finally, the accuracy of the imaging genomics model was determined and verified through calibration decision curve analysis and AUC. The sensitivity was 86.1%, the specificity was 17.6%, the positive predictive value PPV was 81.6%, and the negative predictive value NPV was 23.1%. In addition, the applicant constructed a molecular genomics model based on gene molecular data. The sensitivity was 86.1%, the specificity was 84.6%, the positive predictive value PPV was 96.9%, the negative predictive value NPV was 52.4%, and the accuracy was 85.9%. Although the cumulative results show that the proposed models all have a certain predictive ability, further improvement is needed. Summary of the Invention

[0012] To solve the above technical problems, the objective of the present invention is to provide a method for classifying the benign and malignant nature of thyroid nodules based on clinical information, radiomics, and gene detection. This method combines gene detection, radiomics, and basic clinical information, and based on the Chinese population, establishes a diagnostic model for thyroid nodules based on the above combination. It is used to assist clinicians in distinguishing the benign and malignant nature of thyroid nodules and guiding clinical decision-making.

[0013] To achieve the above objective, the present invention adopts the following technical solutions: A method for classifying the benign and malignant nature of thyroid nodules based on clinical information, radiomics, and gene detection, the method comprising the following steps: 1) Obtain the thyroid ultrasound examination images of thyroid tumor patients with clear postoperative pathological diagnosis, and extract the sagittal and coronal plane graphics of the maximum diameter of the tumor; 2) Preprocess the image data, segment the region of interest of the tumor on the ultrasound image, and use the intraclass correlation coefficient to evaluate the internal and internal consistency of the segmentation of the region of interest; after extracting a number of radiomics features, standardize them to eliminate the unit limitation of the data; 3) Select the ultrasound radiomics features of thyroid cancer and benign thyroid nodules, perform dimensionality reduction, further screen the features using the least absolute shrinkage and selection operator, construct radiomics features using multivariate logistic regression analysis, and then use the selection operator to predict thyroid malignant nodules based on the selected features; calculate the radiomics score for each patient; evaluate the prediction efficiency of the radiomics features using the receiver operating characteristic (ROC) curve of the training set and the test set, and finally obtain the radiomics features specific to thyroid cancer with the best prediction efficiency; 4) Using the pathological diagnosis of thyroid nodules as the dependent variable, and including the detection data of radiomics features and molecular omics as continuous variables, the molecular omics data selects at least the detection data of three genes, namely CLDN10, HMGA2, and LANM3; at the same time, combine some clinicopathological features: gender, age, and BRAF V600E mutation status factors, and through the method of SVM modeling, construct a preoperative diagnosis prediction model for thyroid nodules based on radiomics and molecular omics.

[0014] Preferably, in step 2), the Artificial Intelligence Kit is used to preprocess the image data, including picture gray-scale discretization, intensity normalization, and resolution unification; the ITK-SNAP software is used to manually segment the region of interest of the tumor on the ultrasound image.

[0015] Preferably, the features extracted by radiomics in step 2) include gray-level co-occurrence matrix, histogram, format features, gray-level co-occurrence matrix, gray-level run length matrix, and gray-level size zone matrix.

[0016] Preferably, in step 3), correlation test, Mann-Whitney U test and analysis of variance are used for dimensionality reduction.

[0017] Preferably, in step 4), MATLAB 2022 is used as the platform and LibSVM 3.2 language package is used as the basis. Linear SVM model, non-linear SVM model, LDA model and Risk Score evaluation model are respectively used to construct a prediction model.

[0018] Preferably, in the process of model building in step 4), in terms of model optimization, the grid search method is used to select parameters for the non-linear SVM model. When screening indicators for the constructed model, threshold method, enumeration method, backward method and forward method algorithms are respectively used for model building, and the indicator screening algorithm is improved according to the characteristics of these algorithms.

[0019] Preferably, the SVM model in step 4) is as follows: Plabel = sgn(∑ni = 0 wi exp(−gamma|(xi - x)|2 + b)). The specific parameters of the model are as follows: Parameter Value Parameters 5×1 double nr_class 2 totalSV 27 rho 0.0760 Label 2×1 double nSV 2×1 double sv_coef 27×1 double SVs 27×6 double Preferably, in step 4), for the detection of three genes, CLDN10, HMGA2 and LAMB3, TRIZOL and RevertAid RT kit are used to isolate RNA and reverse transcribe RNA; subsequently, three real-time quantitative PCR analyses are carried out on the ABI prism 7500 sequence detection system using THUNDERBIRD SYBR qPCR-Mix; The primer sequences are as follows: CLDN10: 5'-GAGCTCCGGATAAAGCCAAAG-3' (forward) and 5'-AACAGCGGCTCTACTTCAT-3' (reverse); HMGA2: 5'-ACCCAGGGGAAGCCCAAA-3' (forward) and 5'-CCCTCTGGCCGTTTTTCTCCA-3' (reverse); LAMB3: 5'-GCAGCCTCACAACTACTACAG-3' (forward) and 5'-CCAGGTCTTACCGAAGTCTGA-3' (reverse).

[0020] Furthermore, the present invention also discloses a computer device, including a memory, a processor and a computer program stored on the memory, and the processor executes the computer program to implement the method.

[0021] Furthermore, the present invention also discloses a computer-readable storage medium, on which a computer program or instruction is stored, and when the computer program or instruction is executed by a processor, the method is implemented.

[0022] Furthermore, the present invention also discloses a computer program product, including a computer program or instruction, and when the computer program or instruction is executed by a processor, the method is implemented.

[0023] Due to the above technical solution adopted by the present invention, using the SVM statistical model, a diagnostic combination combining four genes (BRAF, CLDN10, HMGA2, and LANMB3), radiomics, and some basic clinical information (age, gender) is established. The results show that this combination has good discrimination ability, and 96.7% of thyroid nodule samples are correctly classified. The sensitivity is 100%, the specificity is 93.8%, the PPV is 93.3%, and the NPV is 100%. The diagnostic ability of this combination is very good, and the economic cost is greatly reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 . Flow chart of the research design and implementation of the present invention.

[0025] Figure 2 . Expression diagrams of CLDN10, HMGA2, and LANMB3 genes in PTC.

[0026] Figure 3 . Flow chart of the clinical application practice of the thyroid nodule diagnosis model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] The following combines the embodiments of the present invention, and clearly and completely describes the technical solutions in the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present invention.

[0028] 1. Experimental method: 1.1 Specimen collection The samples taken during puncture or surgery are quickly frozen in liquid nitrogen and stored in a refrigerator at a temperature of -80°C. The histopathological sections of each case are retrospectively reviewed by three pathologists with at least 5 years of experience to verify the histological diagnosis and ensure sufficient cancer in the tumor. All patients signed an informed consent form allowing the scientific use of their biological waste materials and information within the scope permitted by law.

[0029] 1.2 RNA extraction and RT-qPCR TRIZOL and RevertAid RT kits were used to isolate RNA and reverse transcribe RNA. Subsequently, three real-time quantitative PCR (RT-qPCR) analyses were performed on an ABI prism 7500 sequence detection system using THUNDERBIRD SYBR qPCR-Mix. The primer sequences were as follows: CLDN10: 5'-GAGCTCCGGATAAAGCCAAAG-3' (forward) and 5'-AACAGCGGCTCTACTTCAT-3' (reverse), HMGA2: 5'-ACCCAGGGGAAGCCCAAA-3' (forward), 5'-CCCTCTGGCCGTTTTTCTCCA-3' (reverse) LAMB3: 5'-GCAGCCTCACAACTACTACAG-3' (forward) and 5'-CCAGGTCTTACCGAAGTCTGA-3' (reverse).

[0030] 1.3 Ultrasonography Preoperative ultrasonography was performed on all patients. Ultrasonography was performed using a GE LOGIQ E9 ultrasound system. During the examination, each patient lay supine on the examination table with the neck fully extended, and the patient was asked to breathe calmly while the thyroid gland and its surrounding lymph node tissues were examined. The number and size, boundary, internal structure, internal echo, and degree of calcification of thyroid nodules were recorded, and the TI-RADS classification was evaluated. Finally, abnormal thyroid nodules and lymph nodes were recorded and reported.

[0031] 1.4 Radiomics data extraction and model construction All patients with pathologically confirmed thyroid tumors after surgery underwent standardized thyroid ultrasonography before surgery. The sagittal and coronal plane images of the maximum diameter of the tumor were extracted, and the image data were preprocessed using the Artificial Intelligence Kit (3.0), including grayscale discretization, intensity normalization, and resolution unification of the pictures. The region of interest (ROI) of the tumor was manually segmented on the ultrasound images using the ITK-SNAP software (www.itksnap.org), and the intraclass correlation coefficient was used to evaluate the internal and internal consistency of the ROI segmentation. When evaluating the consistency, the evaluator needed to perform 2 ROI segmentations for each case within 2 months to evaluate the consistency of the 2 ROI measurement values. The features extracted by radiomics included gray-level co-occurrence matrix, histogram, format features, gray-level co-occurrence matrix, gray-level run length matrix, and gray-level size zone matrix, etc. For each patient, 1131 features were extracted and then standardized to eliminate the unit limitation of the data. Select the ultrasound imaging features of thyroid cancer and benign thyroid nodules, perform dimensionality reduction using relevant tests, Mann-Whitney U test, and analysis of variance, further screen the features using the least absolute shrinkage and selection operator (LASSO), construct radiomics features using multivariate logistic regression analysis, and then use LASSO to predict thyroid malignant nodules based on the selected features. Subsequently, calculate the radiomics score (rad-score) for each patient. This formula is derived from the training set data and then used to determine the rad score of the patient in the test set. Evaluate the prediction efficiency of the radiomics features using the receiver operating characteristic (ROC) curves of the training set and test set, and finally obtain the thyroid cancer-specific imaging features with the best prediction efficiency.

[0032] For the obtained thyroid cancer-specific imaging features, use the MATLAB 2022 platform, based on the LibSVM 3.2 language package (https: / / www.csie.ntu.edu.tw / ~cjlin / libsvm / ), and through the method of SVM modeling, perform mathematical modeling on the previously screened imaging features to construct a preoperative diagnosis model for thyroid nodules based on imaging genomics. In the process of model optimization, for the nonlinear SVM model, use the grid search method to select parameters. When screening the indicators of the constructed model, use algorithms such as the threshold method, enumeration method, backward method, and forward method for modeling respectively, and make necessary improvements to the indicator screening algorithm according to the characteristics of these common algorithms. In the evaluation of the model effect, we mainly use indicators such as accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and AUC, and take accuracy as the most important criterion.

[0033] 1.5 Construction of a diagnostic model for benign and malignant thyroid nodules: In this study, the pathological diagnosis (benign, malignant) of thyroid nodules was used as the dependent variable (0: benign, 1: malignant), and the data of radiomics and molecularomics were used as continuous variables. At the same time, some clinicopathological features were combined: gender (0: female, 1: male), age (continuous variable), BRAF V600E mutation status (1: non-mutated, 1: mutated), etc. A preoperative diagnosis system for thyroid nodules was constructed based on radiomics and molecularomics. In this study, the SVM modeling method was mainly used. Using MATLAB 2022 as the platform and the LibSVM 3.2 language package as the basis, linear SVM models, non-linear SVM models, LDA models, and Risk Score evaluation models were used to construct prediction models. In terms of model optimization, the grid search method was used to select parameters for the non-linear SVM model. When screening indicators for the constructed models, algorithms such as the threshold method, enumeration method, backward method, and forward method (the above four algorithms are not ranked in order) were used for modeling, and the indicator screening algorithm was improved as necessary according to the characteristics of these common algorithms. In the evaluation of the model's effect, we mainly used indicators such as accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and AUC, and accuracy was the most important criterion.

[0034] 1.6 Statistical analysis SPSS 25.0 was used for statistical analysis. Normally distributed data (data conforming to the normal distribution) were presented as mean ± standard deviation (SD) and compared with tests, chi-square analysis, or Fisher's exact test as appropriate. P-values meeting the following conditions were considered statistically significant: 1) all P-values were two-sided, 2) P-value < 0.05. Finally, values including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated to evaluate the diagnostic performance of these tests.

[0035] Results: 2.1 Research flow chart: Our research design flow chart is as Figure 1As shown in the figure. The training set (16 malignant nodules and 12 benign nodules) and validation set (72 malignant nodules, 17 benign nodules) of radiomics included a total of 117 patients, which were used to establish a radiomics model to distinguish benign and malignant thyroid nodules. In addition, we screened 3 candidate differential genes (CLDN10, HMGA2, LAMB3) through transcriptome sequencing analysis and TCGA database analysis, and included some clinically recognized clinicopathological information and BRAF gene. Based on these gene molecular data, we constructed a molecularomics model. Considering that the effects of the above two models were not very ideal, in order to establish a more ideal final combined diagnostic model (C-thyroid model), we additionally enrolled additional training set patients (14 malignant nodules and 16 benign nodules) and additional validation set patients (72 malignant nodules and 13 benign nodules). Finally, it was found that the C-thyroid model had an ideal effect.

[0036] Figure 1 Note: First, we constructed a diagnostic model based on radiomics (radiomics model) using the training set consisting of 16 malignant nodules and 12 benign nodules, and then selected the validation set consisting of 72 malignant nodules and 17 benign nodules to verify it. Second, we screened 3 candidate differential genes (CLDN10, HMGA2, LAMB3) through transcriptome sequencing analysis and TCGA database analysis, and integrated the BRAF gene and some clinical basic information (age, gender). Based on these gene molecular data, we constructed a molecularomics model. Considering that the effects of the above two models (radiomics model and molecularomics model) were still not very ideal, we integrated and optimized the two models, and finally constructed a combined diagnostic model (C-thyroid model) based on radiomics, gene detection (CLDN10, HMGA2, LAMB3, BRAF) and clinical basic information (age, gender). Finally, we used the additional training set consisting of 72 malignant nodules and 13 benign nodules to verify the model, and found that the accuracy, sensitivity, specificity, positive predictive value and negative predictive value of the model were all high.

[0037] 2.2 Establishment of the PTC Radiomics Diagnostic Model 117 patients were divided into a training group and a validation group. The data of the training set were used to screen features and construct a radiomics model, and the validation set was used for validation. Multiple feature selections were performed to identify reliable features and reduce redundancy. Based on the imaging features selected from the training set, six algorithms were used to construct a radiomics model for prediction: KNN, LR, DT, linear support vector machine, Gaussian SVM, and polynomial SVM. Finally, the accuracy of the radiomics model was determined and verified by calibration decision curve analysis and AUC. The cumulative results showed that the proposed model had certain predictive ability but needed to be further improved. The sensitivity was 86.1%, the specificity was 17.6%, the positive predictive value PPV was 81.6%, the negative predictive value NPV was 23.1%, and the accuracy was 73.0% (Table 1).

[0038] Table 1: Diagnostic ability of the radiomics model in the validation set 2.3 Screening and identification of CLDN10, HMGA2, and LAMB3 in 19 pairs of PTCs by whole-transcriptome sequencing To screen for differentially expressed genes, we performed whole-transcriptome sequencing on 19 pairs of PTC tissue samples and identified 39 differentially expressed genes (log 2 FC > 5.0, T count > 1000, P value < 0.01). Combining PubMed literature retrieval and analysis, we screened for significantly differentially expressed genes that intersected with adenocarcinoma again. The results showed that CLDN10, HMGA2, and LANMB3 showed the best diagnostic value and were finally selected. Among the 19 pairs of sequencing data, all three genes showed significant differential expression between tumor and normal tissues ( Figure 2 B). To further confirm the expression of the three genes in PTC, we further analyzed the mRNA expression of these three genes in 496 PTCs and 56 normal tissues in the TCGA database. The results showed that the three genes CLDN10, HMGA2, and LANM3 were highly expressed compared with normal tissues ( Figure 2 A). At the same time, we also verified in 38 pairs of FNA samples ( Figure 2 C, 25 pairs of malignant and 13 pairs of benign) and found that the three genes were also highly expressed. Previous studies have shown that the BRAF gene and some clinical information also have the ability to distinguish between benign and malignant tumors. Therefore, the BRAF gene and some basic clinical information were also included in our study.

[0039] Based on these gene molecular data, we constructed a molecularomics model. The sensitivity was 86.1%, the specificity was 84.6%, the positive predictive value PPV was 96.9%, the negative predictive value NPV was 52.4%, and the accuracy was 85.9% (Table 2).

[0040] Figure 2 Description: Figure 2 Example A shows that in the TCGA database (496 cases of papillary thyroid carcinoma and 56 cases of normal tissues), the mRNA expression levels of three genes, CLDN10, HMGA2, and LANM3, in papillary thyroid carcinoma tissues are significantly higher than those in normal tissues. Figure 2 Example B shows that in the 19 pairs of transcriptome sequencing data from our center, the mRNA expression levels of three genes, CLDN10, HMGA2, and LANM3, in papillary thyroid carcinoma tissues are significantly higher than those in normal tissues. Figure 2 Example C shows that we performed PCR detection on 25 pairs of punctured malignant specimens and 13 pairs of punctured benign specimens. The results show that the mRNA expression levels of three genes, CLDN10, HMGA2, and LANM3, in punctured malignant specimens are significantly higher than those in punctured benign specimens.

[0041] Table 2: Diagnostic ability of the molecular omics model in the validation set 2.4 Establishment of the SVM diagnostic model Through the acquisition of radiomics data and the assignment of variables, a unified database was established after assigning 30 training set variables. Based on the MATLAB 2022 platform and the LibSVM 3.2 language package, the optimal SVM diagnostic model for thyroid nodule diagnosis was established.

[0042] The model was debugged using the C-SVC, RBF kernel function, and grid search method. The grid C bound, grid C step, grid g bound, and grid C step were set to -8 to 8, 0.5, -8 to 9, and 0.5 respectively, and the multiplier for cross-validation was set to 5. Patients were divided into two groups, with positive values indicating malignancy and negative values indicating benignity. The specific parameters of the model are shown in Table 3.

[0043] SVM model: Plabel = sgn(∑ni=0 wi exp(−gamma|(xi - x)|2 + b)) Table 3. Specific parameters of the model Parameter Value Parameters 5×1 double nr_class 2 totalSV 27 rho 0.0760 Label 2×1 double nSV 2×1 double sv_coef 27×1 double SVs 27×6 double 2.5 Training set: Construction of the diagnostic model Given that the accuracy of the above two models is not very satisfactory, we plan to integrate and optimize the above two models and propose a combined diagnostic model. RT-qPCR was used to evaluate the mRNA expression levels of these three genes in 30 training set samples, and the BRAF gene mutation status, clinical basic information, and radiomics information were obtained for constructing the diagnostic model. Using the SVM statistical model, we successfully established a diagnostic combination (C-thyroid model) that combines four genes (BRAF, CLDN10, HMGA2, and LANMB3), radiomics, and some clinical basic information (age, gender). The results showed that this combination had good discriminative ability, and 96.7% of thyroid nodule samples were correctly classified. The sensitivity was 100%, the specificity was 93.8%, the PPV was 93.3%, and the NPV was 100% (Table 4).

[0044] 2.6 Validation set: Validating the diagnostic model To validate the performance of this combination (C-thyroid model), RT-qPCR was used to measure the mRNA expression levels of these three genes in another 85 samples, and the BRAF gene mutation status, clinical information, and radiomics information were obtained. This combination also achieved an accuracy of 97.6% on the validation set. The sensitivity remained 100%, the specificity was 84.6%, the PPV was 97.3%, and the NPV was 100% (Table 1). The results showed that the diagnostic ability of this combination was very good and the economic cost was greatly reduced. The specific process for clinical practice is as Figure 3 shown.

[0045] Table 4: Diagnostic ability of the combined diagnostic model (C-thyroid model) in the training set and validation set Table 4 Note: The accuracy of the combined diagnostic model (C-thyroid model) in the training set was 96.7%, the sensitivity was 100%, the specificity was 93.8%, the positive predictive value was 93.3%, and the negative predictive value was 100%, reaching a relatively high level. Further verification through the validation set found that the validation accuracy of the thyroid nodule diagnostic model was 97.6%, the sensitivity was 100%, the specificity was 84.6%, the positive predictive value was 100%, and the negative predictive value was 100%, still at a relatively ideal level.

[0046] Figure 3 Note: Figure 3The specific practice process planned to be applied in the later clinical stage for the C-thyroid model. First, patients with suspected thyroid nodules need to undergo thyroid ultrasound examination and thyroid fine needle aspiration biopsy (according to clinical guidelines). For patients who are not diagnosed, further radiomics analysis and gene testing are performed. Then, the obtained data is input into the thyroid nodule diagnosis model, and the malignancy of the nodule is evaluated through model calculation.

[0047] The above is the description of the embodiments of the present invention. Through the above description of the disclosed embodiments, those skilled in the art can implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather will conform to the widest scope consistent with the principles and novel points disclosed herein.

Claims

1. A method for classifying benign and malignant thyroid nodules based on clinical information, radiomics, and genetic testing, characterized in that, The method includes the following steps: 1) Obtain thyroid ultrasound examination images of patients with thyroid tumors with clear postoperative pathological diagnoses, and extract the sagittal and coronal plane graphics of the maximum diameter of the tumors; 2) Preprocess the image data, segment the region of interest (ROI) of the tumors on the ultrasound images, and use the intraclass correlation coefficient to evaluate the internal and inter - internal consistency of the ROI segmentation; after extracting several radiomics features, standardize them to eliminate the unit limitations of the data; 3) Select the ultrasound radiomics features of thyroid cancer and thyroid benign nodules, perform dimensionality reduction, further screen the features using the least absolute shrinkage and selection operator, construct radiomics features using multivariate logistic regression analysis, and then use the selection operator to predict thyroid malignant nodules based on the selected features; calculate the radiomics score for each patient; evaluate the prediction efficiency of the radiomics features using the receiver operating characteristic (ROC) curve of the training set and the test set, and finally obtain the radiomics features specific to thyroid cancer with the best prediction efficiency; 4) Take the pathological diagnosis of thyroid nodules as the dependent variable, use the detection data of radiomics features and molecular omics as continuous variables, and the molecular omics data selects at least the detection data of three genes, namely CLDN10, HMGA2, and LANM3; at the same time, combine some clinicopathological features: gender, age, and BRAF V600E mutation status factors, and construct a preoperative diagnosis prediction model for thyroid nodules based on radiomics and molecular omics through the SVM modeling method.

2. The construction method according to claim 1, wherein In step 2), use the Artificial Intelligence Kit to preprocess the image data, including picture gray - level discretization, intensity normalization, and resolution unification; use the ITK - SNAP software to manually segment the ROI of the tumors on the ultrasound images; preferably, the features extracted by radiomics in step 2) include gray - level co - occurrence matrix, histogram, format features, gray - level co - occurrence matrix, gray - level run - length matrix, and gray - level size zone matrix.

3. The construction method according to claim 1, characterized in that, In step 3), use correlation test, Mann - Whitney U test, and variance analysis for dimensionality reduction.

4. The construction method according to claim 1, characterized in that, In step 4), use MATLAB 2022 as the platform and the LibSVM 3.2 language package as the basis, and use the linear SVM model, nonlinear SVM model, LDA model, and Risk Score evaluation model to construct the prediction model respectively.

5. The construction method according to claim 1, characterized in that In step 4), during the model optimization in the modeling process, for the nonlinear SVM model, use the grid search method to select parameters. When screening the indicators of the constructed model, use the threshold method, enumeration method, backward method, and forward method algorithms for modeling respectively, and improve the indicator screening algorithm according to the characteristics of these algorithms.

6. The construction method according to claim 1, wherein In step 4), use the C - SVC, RBF kernel function, and grid search method to debug the model. The grid C boundary, grid C step, grid g boundary, and grid C step are set to - 8 to 8, 0.5, - 8 to 9, and 0.5 respectively, and the multiplier of cross - validation is set to 5; the patients are divided into two groups, with positive values indicating malignancy and negative values indicating benignancy.

7. The construction method according to claim 1, wherein Step 4) The SVM model is as follows: Plabel = sgn(∑ni=0 wi exp(−gamma |(xi - x)|2 + b)); The specific parameters of the model are as follows: 。 8. The construction method according to claim 1, characterized in that, For the detection of three genes, CLDN10, HMGA2, and LAMB3, TRIZOL and RevertAid RT kits were used to isolate RNA and reverse transcribe RNA; subsequently, three real-time quantitative PCR analyses were performed using THUNDERBIRD SYBR qPCR-Mix on an ABI prism 7500 sequence detection system; The primer sequences are as follows: CLDN10: 5'-GAGCTCCGGATAAAGCCAAAG-3' (forward) and 5'-AACAGCGGCTCTACTTCAT-3' (reverse); HMGA2: 5'-ACCCAGGGGAAGCCCAAA-3' (forward) and 5'-CCCTCTGGCCGTTTTTCTCCA-3' (reverse); LAMB3: 5'-GCAGCCTCACAACTACTACAG-3' (forward) and 5'-CCAGGTCTTACCGAAGTCTGA-3' (reverse).

9. A computer device, comprising a memory, a processor, and a computer program stored on the memory, characterized in that, The processor executes the computer program to implement the method described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instruction is executed by the processor, the method described in any one of claims 1-8 is implemented.

Images