Study On The Application Of Biphasic Contrast-enhanced CT Images Based Radiomics In Renal Tumors

PART Ⅰ Study of Radiomics Based Differentiation of Benign and Malignant Small Renal MassesPurpose:To investigate diagnostic performance of radiomic features in differentiation of benign and malignant small renal masses(SRMs)based on biphasic contrast-enhanced computed tomography(CECT)images.Materials and methods:This two-institution retrospective study used biphasic CECT imaage datasets of SRMs(corticomedullary and nephrographic).SRMs were divided into the following three categories:category A,typical angiomyolipoma(AML)with visible fat;category B,benign SRMs without visible fat,including lipid-poor AMLs,oncocytoma,and other rare benign renal tumors;category C,malignant renal tumors.422 SRMs were enrolled at two study sites from August 2009 to December 2017.Radiomics models training was performed by using 350 SRMs datasets obtained from August 2009 to December 2015.The radiomics models were tested with the test group obtained from January 2016 to December 2017.Diagnostic performance of radiomics models in the training and test groups to differentiate benign from malignant SRMs was analyzed by the receiver operating characteristic(ROC)curve.We also compared the specificity and sensitivity of radiomics model with the performance of three experienced radiologists in the test group.Results:For differentiation of categories A and B from category C SRMs in the training group,model corticomedullary,model nephrographic and model biphasic had the accuracies of 84.3%(295/350),84.0%(294/350),and 86.3%(302/350),sensitivities of 81.6%(142/174),89.7%(156/174)and 90.2%(157/174),and specificities of 86.9%(153/176),78.4%(138/176),and 82.4%(145/176),respectively.In the test group,for differentiating categories A and B from category C SRMs,model corticomedullary,model nephrographic and model biphasic had the accuracies of 90.2%(65/72),83.3%(60/72),and 88.9%(64/72),and specificities of 91.2%(31/34),94.1%(32/34),and 94.1%(32/34),respectively;for differentiating category B form category C SRMs,the model corticomedullary,model nephrographic and biphasic had the accuracies of 86%(43/50),78%(39/50),and 84%(42/50),sensitivities of 89.5%(34/38),73.7%(28/38),and 84.2%(32/38),and specificities of 75%(9/12),91.7%(11/12),and 83.3%(10/12),respectively.Compared to radiologist interpretation,model biphasic achieved a similar sensitivity(84.2%versus 84.2%,P=1.000)but higher specificity(83.3%versus 33.3%,p=0.013)in differentiation of category B from category C SRMs in the test group.Conclusion:The radiomics models allow for highly accurate characterization of SRMs by using biphasic contrast-enhanced CT images of the kidney,and when compared to experienced radiologists,model biphasic demonstrated similar sensitivity and improved specificity in differentiation of benign SRMs without visible fat and malignant SRMs.Part Ⅱ Study of Radiomics Based Differentiation of Common Renal Cell Carcinoma SubtypesPurpose:To evaluate diagnostic performance of radiomic features in differentiation of common renal cell carcinoma(RCC)subtypes based on biphasic contrast-enhanced computed tomography(CECT)images.Materials and methods:This single-center retrospective study used biphasic CECT image datasets of renal cell carcinomas from August 2009 to December 2017,including corticomedullary and nephrographic phases images.The pathology confirmed RCCs included 372 clear cell renal cell carcinomas(RCC)and 143 non-clear cell renal cell carcinomas(non-ccRCC)[80 papillary renal cell carcinomas(pRCC)and 63 chromophobe renal cell carcinomas(ChRCC)].Radiomics models training was performed by using 417 RCCs datasets obtained from August 2009 to December 2015.The radiomics models were tested with the test group obtained from January 2016 to December 2017.Diagnostic performance of radiomics models in the training and test groups to differentiate common RCC subtypes was analyzed by the receiver operating characteristic(ROC)curve.Results:In the training group,the classifier area of radiomics models under the ROC curve for discriminating ccRCC and non-ccRCC subtypes by using model corticomedullary,model nephrographic,and model biphasic was 0.878[95%confidential interval(CI),0.843-0.908;accuracy,80.3%(335/417);sensitivity,78.5%(237/302);specificity,85.2%(98/115)],0.783[95%CI,0.740-0.821;accuracy,76.5%(319/417);sensitivity,78.1%(235/302);specificity,73%(84/115)],and 0.877[95%CI,0.842-0.907;accuracy,79.3%(331/417);sensitivity,76.2%(230/302);specificity,87.8%(101/115)],respectively.In the test group,the classifier area of radiomics models under the ROC curve for discriminating ccRCC and non-ccRCC subtypes by using model corticomedullary,model nephrographic,and model biphasic was 0.908[95%CI,0.832-0.957;accuracy,79.6%(78/98);sensitivity,75.7%(53/70);specificity,89.3%(25/28)],0.868[95%CI,0.785-0.928;accuracy,79.6%(78/98);sensitivity,77.1%(54/70);specificity,85.7%(24/28)],and 0.912[95%CI,0.837-0.960;accuracy,82.7%(81/98);sensitivity,80%(56/70);specificity,89.3%(25/28)],respectivelyConclusion:The radiomics models allow for accurate differentiation of common RCC subtypes by using biphasic CECT images of the kidney.