Preliminary Study On The Differentiation Of Benign And Malignant Primary Bone Tumors Based On Preoperative CT And Machine Learning Radiomics

Objectives:Based on preoperative CT scan images of bone tumor patients,the radiomics features containing the identification information of benign and malignant primary bone tumors were extracted and analyzed,and the radiomics models constructed by different ML algorithms for the diagnosis of benign and malignant primary bone tumors were explored.Methods:The pathological examination results of all bone tumor patients retrieved from the database of the pathology department of North Sichuan Medical College Affiliated Hospital were retrospectively analyzed.The data come from January 2016 to October 2021.Excluded metastatic tumors.CT images and medical records of primary bone tumor patients were collected.Finally,312 patients were screened,including 158 benign patients and 154 malignant patients.Using PyRadiomics to extract features from the manually delineated ROIs on the resampled CT images,the training set and validation set were randomly divided according to the 7/3 ratio by python,and feature reproducibility analysis,t-test,and LASSO regression were used to screen out 35 radiomics features that could better identify benign and malignant primary bone tumors on the training set.Using Random Forest Classifier,Support Vector Machine,Logistic Regression Model,K-Nearest Neighbor,Decision Tree,Gradient Booster,Naive Bayes and Ensemble Learning and Simple Artificial Neural Network 9 kinds respectively Machine learning algorithms build classification predictive models.Trained the fitting model on the training set,5-fold cross-validation was used to evaluate model performance,then re-evaluated on the validation set divided separately with the ROC curve.Delong test was performed between the constructed 9 models to determine whether the difference in ROC curves between the models was statistically significant,and to further evaluate the model performance.Results:The results of diagnostic efficacy of primary malignant bone tumors showed that the Light GBM model had the best performance,and its accuracy,specificity,sensitivity,and AUC were 0.96,0.94,0.98,and 0.99,respectively.The differences were statistically significant(P<0.05).The model with the worst performance is Naive Bayes,and its accuracy,specificity,and AUC are only 0.74,0.56,and 0.82,but the Delong test,found that the difference between Naive Bayes and the KNN model is not statistically significant(P>0.05),and the others were statistically significant,indicating that both the KNN model and the Naive Bayes model had a poor diagnostic performance.In terms of sensitivity,the poor TDTC was only 0.80,but the difference between TDTC and SVM,LOG,KNN,and Ensemble was not statistically significant(P>0.05).Conclusions:In this study,based on the patients’ preoperative CT,the radiomics features containing the identification information of primary benign and malignant bone tumor lesions were extracted.The radiomics models constructed by different machine learning algorithms can better predict the benign and malignant bone tumors,with the excellent performance the Light GBM model performed best.This work demonstrates that preoperative CT images can be used to effectively predict benign and malignant primary bone tumors by radiomics methods.