The Application Of CT Radiomics And Machine Learning In The Diagnosis Of Bosniak Ⅲ Renal Masses

Bosniak Ⅲ renal masses cannot be identified as benign or malignant lesions before operation.About 50% of Bosniak Ⅲ renal masses are confirmed as benign after resection,but still,need surgical excision as recommended.Texture analysis refers to the process of extracting texture feature parameters from images and obtaining quantitative or qualitative features through image processing technology.Radiomics automatically collects a large number of image features with high throughput,and extracts and analyzes the effective information from the acquired data,aiming to obtain valuable characteristic parameters and guide the diagnosis and treatment of clinical diseases.Many studies have demonstrated that texture analysis and radiomics are valuable in the diagnosis,treatment evaluation,and prediction of the prognosis,and also able to reflect the information of tissue characteristics objectively,quantitatively,and noninvasively.Machine learning,as a branch of artificial intelligence,has developed into a rapidly growing interdisciplinary approach in recent years.It generates computer algorithms from data and then judges and analyzes new situations through models.Machine learning is widely used in the field of medicine because of its self-learning characteristics and higher accuracy compared to traditional statistics.Based on these,this study had three parts to explore the texture feature parameters with the best diagnostic efficacy,and adopt different dimension reduction methods and machine learning,in order to build models which can accurately predict Bosniak Ⅲ renal masses.These methods may help avoid unnecessary surgical resection and benefit more patients.The three parts are as followed:(1)The application of CT texture analysis in the diagnosis of Bosniak Ⅲ renal masses Objective:To explore the difference between benign and malignant Bosniak Ⅲ renal masses by CT examination with texture analysis.Methods:This retrospective analysis collected 45 patients from June 2012 to November2020 in our hospital,who underwent preoperative CT examinations and had imaging characteristics in line with the Bosniak Ⅲ renal masses.All patients were divided into benign(n=22)and RCC group(n=23).Basic clinical data,preoperative CT images,and postoperative pathological results were obtained.Image J software was used for image preprocessing,and Ma Zda software was used for tumor segmentation and feature extraction to obtain histogram-based,GLCM and GLRLM feature parameters of each period(n=33).For texture parameters with an ICC greater than 0.8,non-parametric Mann-Whitney U test was used to compare the difference between benign and malignant CT texture features,and ROC curves were plotted for texture features with statistical significance.Sensitivity,specificity,cut-off value,Jorden index,and AUC were calculated to analyze and evaluate the diagnostic efficacy of each feature.ROC curves were plotted for the three texture features with the best diagnostic efficacy to evaluate the diagnostic value of multivariate analysis.Results:Among the 33 features extracted,16 features had good inter-observer reliability(ICC>0.8)and significant statistical difference between the two groups(P<0.05).After Holm-Bonferroni correction,8 features were statistically different between the two groups.1.Results of texture analysis in unenhanced CT phase: inverse difference moment,difference entropy,short run emphasis,and long run emphasis showed the statistical differences between benign group and RCC group(P<0.05).Difference entropy had relatively good discriminability and certain accuracy(AUC>0.7).2.Results of texture analysis in corticomedullary phase(CMP): kurtosis,skewness,angular second moment,sum entropy,difference entropy,short run emphasis,and long run emphasis showed the statistical differences between benign group and RCC group(P<0.05).Among them,skewness,sum entropy,difference entropy,and long run emphasis had relatively good discriminability and certain accuracy after Holm-Bonferroni correction(AUC>0.7).3.Results of texture analysis in nephrographic phase(NP): kurtosis,skewness,inverse difference moment,sum entropy,and difference entropy were statistically different between the benign group and the RCC group(P<0.05).Among them,kurtosis,skewness,inverse difference moment,and difference entropy had relatively good discriminability and certain accuracy after Holm-Bonferroni correction(AUC>0.7).4.For the three texture parameters with the best discrimination,an AUC of 0.864 was obtained after the ROC curve was plotted.Multivariate analysis showed a relatively high discriminative ability compared to univariate analysis.Conclusion:CT texture features are correlated to malignancy in Bosniak Ⅲ renal lesions.CTTA may be conducive to differentiating malignant from benign Bosniak Ⅲ renal lesions on CT images.(2)The application of CT texture analysis and machine learning in the diagnosis of Bosniak Ⅲ renal massesObjective:To develop machine learning models based on different algorithms,and to investigate the ability of texture analysis combined with machine learning to differentiate benign from malignant Bosniak Ⅲ renal masses.Methods:This retrospective analysis collected 45 patients from June 2012 to November2020 in our hospital,who underwent preoperative CT examinations and had imaging characteristics in line with the Bosniak Ⅲ renal masses.All patients were divided into benign(n=22)and malignant group(n=23).Basic clinical data,preoperative CT images,and postoperative pathological results were obtained.Image J software was used for image preprocessing,and Ma Zda software was used for tumor segmentation and feature extraction to obtain histogram-based,GLCM and GLRLM feature parameters of each period(n=33).Four dimensionality reduction methods and five machine learning methods were used to reduce the dimension of the texture parameters with the ICC greater than 0.8 and develop models.Sensitivity,specificity,accuracy,and AUC were calculated to evaluate the diagnostic efficacy of each model.Results:A total of 20 groups of machine learning models were built.Among them,the models which used PCA,LASSO and Boruta algorithm as dimension-reduction methods had higher diagnostic efficacy,with the average AUCs of 0.84±0.04,0.86±0.06,and 0.86±0.07,respectively.Among the five machine learning methods,the diagnostic efficiency of the models based on LR and RF algorithm were higher,with the average AUCs of 0.86±0.08 and 0.88±0.04,respectively.In these 20 groups of models,LASSO-LR model(AUC=0.93±0.04,SEN=0.95±0.08,SPE=0.81±0.07,ACC=0.87±0.06),LASSO-RF model(AUC=0.90±0.04,SEN=0.90±0.17,SPE=0.83±0.05,ACC=0.84±0.10),Boruta-SVM model(AUC=0.91±0.07,SEN=0.87±0.23,SPE=0.83±0.21,ACC=0.84±0.21),Boruta-RF model(AUC=0.91±0.08,SEN=0.93±0.13,SPE=0.81±0.02,ACC=0.89±0.10),mRmR-RF model(AUC=0.90±0.02,SEN=0.92±0.07,SPE=0.92±0.08,ACC=0.91±0.04)had significantly better diagnostic efficacy than other models.Conclusion:Machine learning combined with texture analysis can effectively differentiate malignant from benign Bosniak Ⅲ renal lesions and has high diagnostic efficacy,especially in LASSO-LR model,LASSO-RF model,Boruta-SVM model,Boruta-RF model,and mRmR-RF model.The diagnostic efficacy of the models based on texture feature parameters and machine learning was significantly better than the single texture analysis.(3)The application of CT radiomics and machine learning in the diagnosis of Bosniak Ⅲ renal massesObjective:To develop machine learning models based on different algorithms,and to evaluate the ability of radiomics combined with machine learning to differentiate benign from malignant Bosniak Ⅲ renal masses.Methods:This retrospective analysis collected 45 patients from June 2012 to November 2020 in our hospital,who underwent preoperative CT examinations and had imaging characteristics in line with the Bosniak Ⅲ renal masses.All patients were divided into benign(n=22)and malignant group(n=23).Basic clinical data,preoperative CT images,and postoperative pathological results were obtained.Image J software was used for image preprocessing,and Ma Zda software was used for tumor segmentation and feature extraction to obtain feature parameters of each period.A total of 839 radiomics features were obtained.Four dimensionality reduction and five machine learning methods were used to reduce the dimensions of the radiomics features with the ICC>0.8 and develop models.The sensitivity,specificity,accuracy,and AUC were calculated to evaluate the diagnostic efficacy of each model.Results:A total of 20 groups of machine learning models were developed.Among them,model which used LASSO algorithm as dimension-reduction methods had higher diagnostic efficacy,with an average AUC of 0.93±0.04.Among five machine learning methods,model diagnostic efficacy based on SVM,LR and RF algorithm were higher,with the average AUCs of 0.91±0.05,0.90±0.03 and 0.91±0.04,respectively.In these20 groups of models,LASSO-KNN model(AUC=0.95±0.03,SEN=0.88±0.11,SPE=0.92±0.07,ACC=0.89±0.04),LASSO-SVM model(AUC=0.94±0.06,SEN=0.96±0.07,SPE=0.89±0.11,ACC=0.91±0.04),LASSO-LR model(AUC=0.94±0.05,SEN=0.90±0.17,SPE=0.81±0.07,ACC=0.84±0.10),LASSO-RF model(AUC=0.95±0.04,SEN=0.93±0.13,SPE=0.96±0.07,ACC=0.93±0.07),Boruta-SVM model(AUC=0.92±0.03,SEN=0.81±0.06,SPE=0.79±0.06,ACC=0.80±0.03),Boruta-RF model(AUC=0.95±0.03,SEN=0.96±0.07,SPE=0.91±0.16,ACC=0.93±0.07),mRmR-SVM model(AUC=0.94±0.06,SEN=0.94±0.10,SPE=0.85±0.17,ACC=0.89±0.14),mRmR-LR model(AUC=0.91±0.09,SEN=0.83±0.02,SPE=0.80±0.11,ACC=0.81±0.07)had better diagnostic efficiency than other models.Conclusion:Radiomics combined with machine learning can effectively differentiate malignant from benign Bosniak Ⅲ renal lesions and has high diagnostic efficacy,especially LASSO-KNN model,LASSO-SVM model,LASSO-LR model,LASSO-RF model,Boruta-SVM model,Boruta-RF model,mRmR-SVM model,and mRmR-LR model.The diagnostic efficacy of the models based on radiomics parameters and machine learning is significantly better than those based on texture analysis alone and texture analysis combined with machine learning.