Thyroid cancer early diagnosis model based on improved deep forest
Through ReliefF feature selection and improved deep random forest algorithm, combined with the cascaded forest structure of LightGBM and XGBoost, the problems of feature loss and low diagnostic efficiency in early diagnosis of thyroid cancer are solved, and efficient laboratory blood detection assisted diagnosis is achieved.
Patent Information
- Application Number
- CN202510581123.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-07
- Publication Date
- 2025-08-15
- Estimated Expiration
- Not applicable · inactive patent
Smart Images
Figure CN120496803A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for early diagnosis of thyroid cancer, and in particular to an early diagnosis model method for thyroid cancer based on an improved deep forest. Background Art
[0002] Thyroid cancer, a malignant disease, has shown an increasing incidence in most countries and regions worldwide in recent years. In the early stages of thyroid cancer, cure rates are relatively high and treatment outcomes are generally good. However, once the disease progresses, the patient's risk of death increases significantly, and the cost of treatment also rises significantly. Currently, a variety of methods are available for diagnosing thyroid cancer, including ultrasound, laboratory tests, imaging, and fine-needle aspiration biopsy. However, each of these methods has its limitations. For example, there is no globally unified standard for laboratory serum testing, and some indicators lack sufficient specificity. Although fine-needle aspiration biopsy is considered the gold standard for diagnosis, it is invasive and expensive, making it difficult for patients to undergo.
[0003] Currently, in the field of early diagnosis of thyroid cancer, modeling research on laboratory blood tests mostly focuses on logistic regression models. In contrast, research using machine learning and deep learning technologies is relatively rare. In addition, the indicators selected in these studies are mostly limited to thyroid function-related parameters, such as thyroglobulin (TG), thyroglobulin antibodies (TGAb), and thyroid-stimulating hormone (TSH). However, these indicators lack specificity and may lead to the neglect of some indicators with good diagnostic value for papillary thyroid cancer (PTC). Therefore, based on machine learning, feature screening of blood indicators, finding the optimal model, and using computer-aided diagnosis (CAD) to diagnose thyroid cancer early can not only improve diagnostic efficiency but also provide doctors with more objective and accurate diagnostic results, thus having important clinical application value. Summary of the Invention
[0004] In view of the limitations of the existing technology, the purpose of this invention is to propose an early diagnosis model method for thyroid cancer based on an improved deep forest. First, the ReliefF feature selection method is used to identify discriminative features, and then the improved deep random forest algorithm is used to train the model for the selected features, thereby effectively pre-diagnosing thyroid cancer through laboratory blood tests.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing an early diagnosis model for thyroid cancer based on an improved deep forest, comprising the following steps:
[0006] Step 1: Construct a thyroid patient dataset. Based on the RW-FMEA strategy, clinical data were collected. Patients diagnosed with benign thyroid nodules and thyroid cancer were selected as research subjects. The former were defined as the benign group, and the latter as the malignant group for subsequent analysis.
[0007] Step 2: Feature extraction from various laboratory blood tests including patient basic information, thyroid function indicators, tumor indicators, and biochemical indicators;
[0008] Step 3: Data preprocessing. The extracted features are first cleaned and then normalized using Min-Max Scaling to scale the features in each sample to between 0 and 1. The data is then divided into training data and test data. The training set is used to train the model, while the test set is used to evaluate the model's performance.
[0009] Step 4: Before model training and prediction, feature variables must be screened. Therefore, we use the Relief feature selection method to identify discriminative features on the processed data. The weight of each feature is calculated according to Formula 1.
[0010]
[0011] Perform calculations;
[0012] Step 5: Build a model classifier based on different trees;
[0013] Step 6: Obtain the classifier. In order to better evaluate the performance of the classifier, the accuracy, specificity and sensitivity are used to evaluate the performance of the classifier, as shown in formula (2), formula (3) and formula (4) respectively:
[0014]
[0015]
[0016] Through training, when the three criteria of accuracy, specificity and sensitivity of the classifier all reach the set threshold, it is considered that a classifier model with satisfactory performance is obtained; the trained classifier model is used to classify benign and malignant tumors of the thyroid disease to be diagnosed to assist doctors in diagnosis.
[0017] As a preferred technical solution of the present invention, the method for selecting the research object in step 1 is:
[0018] Step 1-1: Inclusion criteria: Patients must be adults (aged 18 or above) with complete clinical data; have no history of thyroid surgery or treatment; and have no other tumors or medical history except benign thyroid nodules or papillary thyroid carcinoma.
[0019] Step 1-2: Exclusion criteria: Minor patients under the age of 18; patients with serious missing clinical data; patients with a history of other tumors; patients with mental illness and major organ dysfunction are excluded.
[0020] As a preferred technical solution of the present invention, in the formula 1 applied in step 4, diff(A, R1, R2) represents the difference between samples R1 and R2 in feature A, M j (C) indicates class The jth nearest neighbor sample in .
[0021] As a preferred technical solution of the present invention, the method for constructing the model classifiers of different trees in step 5 is:
[0022] The selected features are input into the cascade forest of the deep forest algorithm for training. Another focus of the present invention is to integrate three different tree-based classifiers at each level of the deep forest model, utilizing the advantages of different trees to enhance classification performance.
[0023] The cascaded structure of a deep forest is similar to a neural network, but instead of neurons and weights, it uses decision trees and random forests as its basic units. Each cascade layer takes the output of the previous layer as input and outputs the processed information to the next layer. This layer-by-layer processing helps gradually build higher-order features, thereby improving the model's representational learning capabilities.
[0024] In this step, the integration of three different tree-based classifiers is specifically implemented by the following steps:
[0025] Step 5-1: Introduce LightGBM and XGBoost algorithms to improve the cascade forest structure of the deep forest algorithm; each layer consists of two random forests, LightGBM and XGBoost.
[0026] Step 5-2: Input each feature vector into multiple random forests (including two random forests, LightGBM and XGBoost), each of which outputs a class vector.
[0027] Step 5-3: Concatenate the class vectors output by each forest to obtain the output of this layer. Then, concatenate the output of this layer with the original features or the output of the previous layer as the input of the next layer.
[0028] Step 5-4: Repeat the above process until the iteration termination condition is met.
[0029] As a preferred technical solution of the present invention, in Formula 3, Formula 4 and Formula 5 applied in step 6, P is the number of malignant specimens, N is the number of benign specimens, TP is the number of correctly classified malignant specimens, TN is the number of correctly classified benign specimens, FP is the number of benign specimens misclassified, and FN is the number of misclassified malignant specimens.
[0030] An early diagnosis model for thyroid cancer based on an improved deep forest method first uses the ReliefF feature selection method to identify discriminative features, and then uses the improved deep random forest algorithm to train the model on the selected features, thereby effectively pre-diagnosing thyroid cancer through laboratory blood tests.
[0031] The early diagnosis method of thyroid cancer based on the improved deep forest of the present invention has the following beneficial effects: First, with regard to feature extraction, the ReliefF feature selection method identifies discriminative features, avoiding the problem of feature loss caused by traditional medical statistics; Second, the extracted features are trained on the model using the improved deep forest algorithm. Deep forests are widely used in data classification and retrospective field analysis, but the accuracy of small sample data needs to be improved. The improved deep forest integrates algorithms of multiple trees and utilizes the advantages of different trees to enhance classification performance; Third, laboratory blood tests of patients are used for auxiliary diagnosis in the early stages to achieve early diagnosis and early treatment. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 This is a schematic diagram of the grouping of research objects in the present invention;
[0033] Figure 2 Schematic diagram of the method for constructing a deep forest in the present invention. DETAILED DESCRIPTION
[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0035] The early diagnosis model for thyroid cancer based on the improved deep forest method uses the ReliefF feature selection method to identify discriminative features, and then uses the improved deep random forest algorithm to train the model on the selected features, thereby effectively pre-diagnosing thyroid cancer through laboratory blood tests.
[0036] The method for constructing an early diagnosis model for thyroid cancer based on an improved deep forest includes the following steps:
[0037] Step 1: Construct a thyroid patient dataset, collect clinical data based on the RW-FMEA strategy, select patients diagnosed with benign thyroid nodules and thyroid cancer as research subjects, define the former as the benign group and the latter as the malignant group for subsequent analysis, such as Figure 1 As shown;
[0038] Step 1-1: Inclusion criteria: Patients must be adults (aged 18 or above) with complete clinical data; have no history of thyroid surgery or treatment; and have no other tumors or medical history except benign thyroid nodules or papillary thyroid carcinoma.
[0039] Step 1-2: Exclusion criteria: Minors under the age of 18; patients with severe missing clinical data; patients with a history of other tumors; patients with mental illness and major organ dysfunction;
[0040] Step 2: Feature extraction from various laboratory blood tests including patient basic information, thyroid function indicators, tumor indicators, and biochemical indicators;
[0041] Step 3: Data preprocessing. The extracted features are first cleaned and then normalized using Min-Max Scaling to scale the features in each sample to between 0 and 1. The data is then divided into training data and test data. The training set is used to train the model, while the test set is used to evaluate the model's performance.
[0042] Step 4: Before model training and prediction, feature variables must be screened. Therefore, we use the Relief feature selection method to identify discriminative features on the processed data. The weight of each feature is calculated according to Formula 1.
[0043] Calculate, in the above formula, diff(A,R1,R2) represents the difference between samples R1 and R2 on feature A, M j (C) indicates class The jth nearest neighbor sample in ;
[0044] Step 5: Build a model classifier based on different trees:
[0045] The selected features are input into the cascade forest of the deep forest algorithm for training. Another focus of the present invention is to integrate three different tree-based classifiers at each level of the deep forest model, taking advantage of the advantages of different trees to enhance the classification performance. Figure 2shown.
[0046] The cascaded structure of a deep forest is similar to a neural network, but instead of neurons and weights, it uses decision trees and random forests as its basic units. Each cascade layer takes the output of the previous layer as input and outputs the processed information to the next layer. This layer-by-layer processing helps gradually build higher-order features, thereby improving the model's representational learning capabilities.
[0047] In this step, the integration of three different tree-based classifiers is specifically implemented by the following steps:
[0048] Step 5-1: Introduce LightGBM and XGBoost algorithms to improve the cascade forest structure of the deep forest algorithm; each layer consists of two random forests, LightGBM and XGBoost.
[0049] Step 5-2: Input each feature vector into multiple random forests (including two random forests, LightGBM and XGBoost), each of which outputs a class vector.
[0050] Step 5-3: Concatenate the class vectors output by each forest to obtain the output of this layer. Then, concatenate the output of this layer with the original features or the output of the previous layer as the input of the next layer.
[0051] Step 5-4: Repeat the above process until the iteration termination condition is met;
[0052] Step 6: Obtain the classifier. In order to better evaluate the performance of the classifier, the accuracy, specificity and sensitivity are used to evaluate the performance of the classifier, as shown in formula (2), formula (3) and formula (4) respectively:
[0053]
[0054] In formula 2, formula 3, and formula 4, P is the number of malignant specimens, N is the number of benign specimens, TP is the number of correctly classified malignant specimens, TN is the number of correctly classified benign specimens, FP is the number of misclassified benign specimens, and FN is the number of misclassified malignant specimens;
[0055] Through training, when the three criteria of accuracy, specificity and sensitivity of the classifier all reach the set threshold, it is considered that a classifier model with satisfactory performance is obtained; the trained classifier model is used to classify benign and malignant tumors of the thyroid disease to be diagnosed to assist doctors in diagnosis.
[0056] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. The method for constructing an early diagnosis model for thyroid cancer based on an improved deep forest algorithm includes the following steps: Step 1: Construct a thyroid patient dataset. Based on the RW-FMEA strategy, clinical data were collected. Patients diagnosed with benign thyroid nodules and thyroid cancer were selected as research subjects. The former were defined as the benign group, and the latter as the malignant group for subsequent analysis. Step 2: Feature extraction from various laboratory blood tests including patient basic information, thyroid function indicators, tumor indicators, and biochemical indicators; Step 3: Data preprocessing. The extracted features are first cleaned and then normalized using Min-Max Scaling to scale the features in each sample to between 0 and 1. The data is then divided into training data and test data. The training set is used to train the model, while the test set is used to evaluate the model's performance. Step 4: Before model training and prediction, feature variables must be screened. Therefore, we use the Relief feature selection method to identify discriminative features on the processed data. The weight of each feature is calculated according to Formula 1. Perform calculations; Step 5: Build a model classifier based on different trees; Step 6: Obtain the classifier. In order to better evaluate the performance of the classifier, the accuracy, specificity and sensitivity are used to evaluate the performance of the classifier, as shown in formula (2), formula (3) and formula (4) respectively: Through training, when the three criteria of accuracy, specificity and sensitivity of the classifier all reach the set threshold, it is considered that a classifier model with satisfactory performance is obtained; the trained classifier model is used to classify benign and malignant tumors of the thyroid disease to be diagnosed to assist doctors in diagnosis.
2. The method for constructing an early diagnosis model for thyroid cancer based on an improved deep forest according to claim 1, characterized in that: The method for selecting research subjects in step 1 is: Step 1-1: Inclusion criteria: Patients must be adults (aged 18 or above) with complete clinical data; have no history of thyroid surgery or treatment; and have no other tumors or medical history except benign thyroid nodules or papillary thyroid carcinoma. Step 1-2: Exclusion criteria: Minor patients under the age of 18; patients with serious missing clinical data; patients with a history of other tumors; patients with mental illness and major organ dysfunction are excluded.
3. The method for constructing an early diagnosis model for thyroid cancer based on an improved deep forest according to claim 1, characterized in that: In Formula 1 applied in step 4, diff(A, R1, R2) represents the difference between samples R1 and R2 on feature A, and M j (C) indicates class The jth nearest neighbor sample in .
4. The method for constructing an early diagnosis model for thyroid cancer based on an improved deep forest according to claim 1, characterized in that: The construction method of the model classifier of different trees in step 5 is: The selected features are input into the cascade forest of the deep forest algorithm for training. Another focus of the present invention is to integrate three different tree-based classifiers at each level of the deep forest model, utilizing the advantages of different trees to enhance classification performance, as shown in Figure 2. The cascaded structure of a deep forest is similar to a neural network, but instead of neurons and weights, it uses decision trees and random forests as its basic units. Each cascade layer takes the output of the previous layer as input and outputs the processed information to the next layer. This layer-by-layer processing helps gradually build higher-order features, thereby improving the model's representational learning capabilities. In this step, the integration of three different tree-based classifiers is specifically implemented by the following steps: Step 5-1: Introduce LightGBM and XGBoost algorithms to improve the cascade forest structure of the deep forest algorithm; each layer consists of two random forests, LightGBM and XGBoost. Step 5-2: Input each feature vector into multiple random forests (including two random forests, LightGBM and XGBoost), each of which outputs a class vector. Step 5-3: Concatenate the class vectors output by each forest to obtain the output of this layer. Then, concatenate the output of this layer with the original features or the output of the previous layer as the input of the next layer. Step 5-4: Repeat the above process until the iteration termination condition is met.
5. The method for constructing an early diagnosis model for thyroid cancer based on an improved deep forest according to claim 1, characterized in that: In Formula 3, Formula 4, and Formula 5 applied in step 6, P is the number of malignant specimens, N is the number of benign specimens, TP is the number of correctly classified malignant specimens, TN is the number of correctly classified benign specimens, FP is the number of misclassified benign specimens, and FN is the number of misclassified malignant specimens.
6. An early diagnosis model for thyroid cancer based on an improved deep forest, characterized by: This early diagnosis model for thyroid cancer based on improved deep forests first uses the ReliefF feature selection method to identify discriminative features, and then uses the improved deep random forest algorithm to train the model on the selected features, thereby effectively pre-diagnosing thyroid cancer through laboratory blood tests.