MRI-based Radiomics And Deep Learning For Predicting Benign And Malignant Renal Tumor And Renal Cell Carcinoma Aggressiveness

Background: Renal tumors encompass a broad disease spectrum including benign and indolent lesions to aggressive and invasive malignancies.The increasing use of cross-section imaging has resulted the rising incidental detection of asymptomatic renal tumors.These tumors are often surgically removed because of radiographically suspicious for renal cell carcinoma(RCC).However,approximately 25%of surgically removed renal tumors are reported to be benign.Even these tumors were RCCs,they were diagnosed when smaller in size and lower in stage,leading to consideration of alternative treatment options for lower risk lesions,including percutaneous ablation and active surveillance.Therefore,pre-treatment differentiation benign from malignant renal tumors and assessment of tumor aggressiveness is now of supreme importance for risk stratification and clinical decision making.Purpose: 1.To build an automatic machine learning model based on magnetic resonance imaging(MRI)radiomics to noninvasively predict SSIGN(Stage,Size,Grade and Necrosis)score in patients with clear cell renal cell carcinoma(cc RCC).2.To develop an MRI-based deep learning model to predict the histological grade of patients with RCC in stage I and II.3.To develop MRI-based radiomics and deep learning models to distinguish benign from malignant renal tumors.Methods and material: 1.A multicenter cohort of 364 histopathologically confirmed cc RCC lesions(254 lesions in training set and 110 lesions in test set)were retrospectively identified.79 non-texture and 94 texture features were extracted from the preoperative MRI images.A tree-based pipeline optimization tool(TPOT)was used to build an automatic machine learning model.14 feature selection methods and 10 machine learning classifiers were trained and tested on features from the same splits of patients to select the top-performing manual optimized machine learning model.Accuracy,sensitivity,specificity,area under the precision recall curve(PR AUC)and area under the receiver operating characteristic curve(ROC AUC)were calculated to evaluate the performance of TPOT model and the manual optimized machine learning model on the test set.2.A multicenter cohort of 430 histopathologically confirmed RCC lesions in stage I-II were retrospectively identified.The353 Fuhrman graded RCCs were divided into training,validation and test set with a 7:2:1 split,the 77 WHO/ISUP graded RCCs were used as a separate WHO/ISUP test set.A deep learning model applying the residual convolutional neural network was developed based on preoperative MRI images.Clinical variables were fed into a separate model that used logistic regression for prediction of histological grade.An ensemble model was created using a bagging classifier combining the output of the clinical variable model and MRI-based deep learning models.F1 score,accuracy,sensitivity,specificity,positive predictive value,negative predictive value,PR AUC and ROC AUC were calculated to evaluate the performance of the deep learning model.3.A multicenter cohort of 1162 renal tumors confirmed by histology or imaging were retrospectively identified.655 lesions were RCCs based on histopathology,162 lesions were benign confirmed by histopathology,and 345 benign renal tumors were diagnosed radiographically.The images of 1162 renal tumors were randomly divided into training set(n = 816),validation set(n = 234)and test set(n = 112)in a 7:2:1 split.All 345 radiographically diagnosed lesions were kept within the training or validation set.A deep learning model applying the residual convolutional neural network was developed based on preoperative MRI images.Clinical variables were fed into a separate model that used logistic regression for prediction of malignancy.An ensemble model was created using a bagging classifier combining the output of the clinical variable model and MRI-based deep learning models.At the same time,the TPOT model and manually optimized machine learning model were built separately based on the MRI radiomics features.In addition,four expert radiologists evaluated unsegmented MRI images of the renal lesions in the test set for malignancy.F1 score,accuracy,sensitivity,specificity,positive predictive value,negative predictive value,PR AUC and ROC AUC were calculated to evaluate the performance of the deep learning model and radiomics model compared with expert interpretation.Results: 1.The manually optimized radiomics model using Random Forest classification and Analysis of Variance feature selection method achieved a test accuracy of 0.89(95% CI,0.82–0.94),specificity of 0.95(95% CI,0.88–0.98),sensitivity of 0.72(95% CI,0.54–0.85),ROC AUC of 0.89 and PR AUC of 0.81.The TPOT using Extra Trees Classifier achieved a test accuracy of 0.89(95% CI,0.82–0.94),specificity of 0.95(95% CI,0.88–0.98),sensitivity of 0.72(95% CI,0.54–0.85),ROC AUC of 0.94 and PR AUC of 0.83.2.The ensemble deep learning model achieved a test F1 score of 0.80,accuracy of 0.88(95% CI,0.73–0.96),sensitivity of 0.89(95% CI,0.74–0.96),specificity of 0.88(95% CI,0.73–0.96),ROC AUC of 0.91 and PR AUC of 0.77 in the Fuhrman test set and a test F1 score of 0.79,accuracy of 0.83(95% CI,0.73–0.90),sensitivity of 0.92(95% CI,0.84–0.97),specificity of 0.78(95% CI,0.68–0.86),ROC AUC of 0.94 and PR AUC of 0.89 in the WHO/ISUP test set.3.Compared with all experts averaged,the ensemble deep learning model had higher test accuracy(0.70 vs.0.60,P = 0.053),sensitivity(0.92 vs.0.80,P = 0.017)and specificity(0.41 vs.0.35,P =0.450).Compared with the most optimized radiomics model,the ensemble deep learning model had higher test accuracy(0.70 vs.0.62,P= 0.081),sensitivity(0.92 vs.0.79,P = 0.012)and specificity(0.41 vs.0.39,P = 0.770).Conclusions: 1.Preoperative MRI-based radiomics can accurately predict SSIGN score of cc RCC.The TPOT model performed slightly better than the manually optimized machine learning model.These results suggest MRI-based radiomics may be beneficial for pretreatment risk stratification and personal treatment in patients with cc RCC.2.An MRI-based deep learning model was developed to noninvasively differentiate low from high histological grade in stages I and II RCCs with high accuracy,which can provide valuable information for guiding management and developing appropriate surveillance programs.3.An MRI-based deep learning model was developed to noninvasively classify benign and malignant renal tumors with good accuracy,sensitivity and specificity comparable with experts and radiomics.If further validated,the model can spare patients unnecessary biopsy/surgery and help guide management in a clinical setting.