A method for identifying ground-glass pulmonary nodules

Abstract

The present invention discloses a method for identifying ground-glass pulmonary nodules, including extracting features from the marked pulmonary nodule region; sending the extracted features to a CatBoost model for feature selection, dividing the ground-glass pulmonary nodule feature data set after feature selection into a training set and a test set, and sending them to the first-layer learner in a Stacking ensemble learning model, using 5-fold cross validation on the training set to implement training and testing of each base classifier, and after each base model training and testing is completed, splicing the output results to obtain a new data set, using the data set as the input of the second-layer learner for training and testing, and finally completing the identification of ground-glass pulmonary nodules. The present invention first extracts the original imaging genomics features, selects the optimal features, and then uses the Stacking ensemble learning method, and introduces four learners of random forest, decision tree, KNN, and LightGBM as base classifiers, and SVM as a meta-learner to improve the accuracy of ground-glass pulmonary nodule recognition.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and in particular relates to a feature selection and ensemble learning method for medical images, and is mainly used for the recognition of ground-glass lung nodules. Background Art

[0002] In recent years, with the rapid development of artificial intelligence, computer-aided diagnosis (CAD) technology of medical images has been increasingly used in doctors' diagnosis of lung nodules. CAD diagnostic technology mainly follows the following steps: 1. Segmentation of candidate lung nodules 2. Extraction and optimization of lung nodule features 3. Classification and identification of lung nodules. In step 2, due to different extraction methods, we will obtain multiple types of nodule features. Different types of features have different degrees of influence on the classification of lung nodules. Too many features may be redundant, resulting in higher computational complexity and lower correct recognition rate. Too few features may lead to the loss of important features, which reduces the performance of the machine learning model. In addition, in the selection of lung nodule classifiers, most researchers use a single machine learning method, but for more complex medical images, a single machine learning model still has certain limitations.

[0003] In order to improve the recognition accuracy of ground-glass pulmonary nodules, the present invention proposes a method based on CatBoost feature selection and Stacking ensemble learning for the recognition of ground-glass pulmonary nodules. This method firstly targets the redundancy of high-dimensional features of ground-glass pulmonary nodules, models them through the CatBoost method, obtains the importance of each original data feature, and selects data features with higher importance to construct the optimal feature subset; secondly, in order to solve the problem of low recognition accuracy, four base classifiers, namely random forest, decision tree, KNN, and LightGBM, and SVM are introduced as meta-classifiers in the Stacking ensemble process, which improves the recognition accuracy of ground-glass pulmonary nodules. Summary of the invention

[0004] The purpose of the present invention is to provide a method for identifying ground-glass pulmonary nodules, thereby solving the current problem of low accuracy in identifying ground-glass pulmonary nodules.

[0005] In order to achieve the above objectives, the technical methods adopted are as follows:

[0006] Step 1), firstly, feature extraction is performed on the experimental data, and a total of 45 features are extracted for each lung nodule;

[0007] Step 2), CatBoost feature selection: the 45 radiomics features extracted in step 1) are standardized and modeled using the CatBoost method to obtain the importance of 45-dimensional data features, and 27 features with higher importance are selected to form the optimal feature data set;

[0008] Step 3) construct a Stacking ensemble learning model, use the optimal feature set obtained in step 2) as a data set, input it into the ensemble learning model, and finally obtain the recognition result of ground-glass lung nodules.

[0009] The ground-glass lung nodule recognition method provided by the present invention first sends the extracted lung nodule imaging features to the CatBoost method for modeling to obtain the optimal feature data set after feature selection, and then uses the Stacking ensemble learning model to classify the data set, which can effectively improve the accuracy of ground-glass lung nodule recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a diagram of the radiomics feature extraction process for lung nodules;

[0011] Figure 2 The importance of the 45-dimensional features of the original feature data of lung nodules;

[0012] Figure 3 Schematic diagram of the process of identifying ground-glass lung nodules according to the present invention. DETAILED DESCRIPTION

[0013] In order to enable those skilled in the art to better understand the technical solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0014] According to the embodiment of the present invention, the experimental data set of the present invention adopts the lung CT data set provided by the General Hospital of North China Petroleum Administration Bureau.

[0015] Step 1: See Figure 1Schematic diagram of the invention, according to the coordinate information of the nodule, the nodule area (i.e., the three-dimensional data block) is extracted in the three-dimensional image data, the grayscale values ​​below 60 on the obtained three-dimensional data block of the lung nodule are set to 0, and the imaging omics features on the data block are calculated. For each nodule, 45 features were extracted, including 15 first-order features (energy, statistics energy, kurtosis, maximum, minimum, mean, mean deviation, median, range, root mean squared, skewness, standard deviation, variance, entropy, and uniformity), 20 texture eigenvalues ​​extracted based on the principle of gray-level co-occurrence matrix (GLCM) (autocorrelation, cluster prominence, cluster shade, cluster tendency, contrast, correlation, difference entropy, dissimilarity, glcm energy, glcm entropy, homogeneity1, informational measure of correlation1, informational measure of correlation2, inverse difference normalized, inversedifference moment normalized, inverse variance, maximum probability, sum average, sum entropy, and sum variance), and 10 eigenvalues ​​extracted based on the principle of gray-level run length matrix (GLRLM) (short run emphasis, long run emphasis, run length nonuniformity, run percentage, low gray level run emphasis, high gray level run emphasis, short run gray level emphasis, short run high gray level emphasis, long run low gray level emphasis, long run high gray level emphasis).

[0016] Step 2: Use CatBoost method to build model. The 45-dimensional radiomics features obtained after the above feature extraction are selected using CatBoost method. The loss_function is RMSE, the iterations (maximum number of trees) is 1000, the learning_rate is 0.03, and the depth is 6. The importance of the 45-dimensional features of the original data is obtained. The importance is as follows: Figure 2 As shown in Figure 1, the 27 more important features in the figure constitute a new optimal feature data set.

[0017] Step 3.1: Build a Stacking ensemble learning model. The Stacking model is usually designed as a 2-layer structure, as shown in the figure. Figure 3 . In the first layer, multiple base learners are constructed. The present invention uses four learners, namely LightGBM (Light GradientBoosting Machine), KNN (K-Nearest Neighbor), decision tree (DecisionTree), and random forest (RF), as the base classifiers of the first layer. The LightGBM learner adopts a learning_rate (learning rate) of 0.1 and n_estimators (number of iterations) of 1000. The KNN learner adopts n_neighbors (k value in KNN) of 5, weights (weights of neighboring samples of each sample) of uniform, and algorithm (algorithm used for the radius-limited nearest neighbor method) of auto. The evaluation criterion of the decision tree learner when using criterion (splitting node) is the Gini index, and the strategy splitter when specifying the splitting node is the best optimal splitting strategy. The random forest learner adopts n_estimators (the number of numbers in the forest) of 800 and max_depth (the maximum depth of the number) of 100, and performs a 5-fold cross validation on the above classifiers, which is specifically implemented as the following process.

[0018] First, divide the data obtained after feature selection in step 2 into a training set and a test set. Secondly, divide the training set into 5 parts, denoted as S_1-S_5. First, use the 4 parts of data S_1-S_4 to train the LightGBM model, and use the trained LightGBM model to predict S_5 and the test set; then use S_1-S_3 and S_5 to retrain the LightGBM model, and predict S_4 and the test set. Similarly, after 5-fold cross-validation, the LightGBM model can obtain the prediction values ​​of S_1-S_5, concatenate the above prediction values ​​in their respective order, and obtain the first set of prediction values. The 5 sets of prediction values ​​of the test set are averaged, and the average value is obtained as the first set of prediction values. Similarly, repeat the above steps for the three base models of KNN, decision tree, and random forest of the first-layer learner to obtain the second, third, and fourth groups of prediction values ​​for the training set and the test set. Finally, the four sets of prediction values ​​of the training set output by the first-layer base classifier are concatenated to construct a new training set, and the four sets of prediction values ​​of the test set output by the first-layer base classifier are concatenated to construct a new test set.

[0019] Step 3.2: Construct the second-layer meta-classifier. The present invention uses a support vector machine (SVM) as a classifier, adopts a radial basis function kernel (sigmoid), a regularization coefficient of 1, and a hyperparameter of 0.5. The new training set and test set obtained in step 3.1 are input into the second-layer meta-learner for training to obtain the final classification result.

Claims

1. A method for identifying ground glass pulmonary nodules, characterized by: include: Step 1), extract 45 radiomics features from the three-dimensional block where the lung nodules are located; Step 2) Use the CatBoost method to build a model, take the feature data set extracted in step 1) as input, analyze all input variables, obtain the importance of 45-dimensional features of the original data, and select 27 features that are highly correlated with the classification of ground glass nodules; Step 3) construct a Stacking ensemble learning model, input the filtered feature set obtained in step 2) into the model to complete the identification of ground-glass lung nodules.

2. The method for identifying ground glass pulmonary nodules according to claim 1, characterized in that: The 45 imaging features extracted in step 1) are characterized in that the grayscale values ​​below 60 on the obtained three-dimensional data block of the lung nodule are set to 0, and then the imaging features on the data block are calculated.

3. The method for identifying ground glass pulmonary nodules according to claim 1, characterized in that: The step 2) uses the CatBoost method to build a model, which is characterized in that the CatBoost model is used to select features for the problem of ground glass nodule feature data redundancy; the prediction value change (Prediction Values ​​Change, PVC) provided in CatBoost is an importance measurement indicator, which shows the average change degree of the prediction value when the feature value changes, and CatBoost uses a tree as an algorithm based on the model. By analyzing whether the original 45 imaging genomics features obtained in step 1) are built into the tree, the importance of the variable is judged by comparing the differences. During the experiment, if the model performance changes greatly, it means that the feature is more important, and finally a feature importance ranking set is generated, thereby reducing feature redundancy.

4. The method for identifying ground glass pulmonary nodules according to claim 1, characterized in that: The step 3) constructs a Stacking ensemble learning model, which is characterized by using the Stacking method for integration; the Stacking classification method is an integration strategy that combines multiple single models together, and uses 5-fold cross-validation to train four base classifiers, random forest, decision tree, KNN, and LightGBM, and constructs the predicted values ​​obtained after training into a new data set as the input of the second-layer meta-learner SVM, and the meta-learner outputs the results to finally complete the classification of ground glass nodules.

Images