The Application Of Magnetic Resonance Imaging Based Radiomic Nomogram For Discrimination Between Benign And Malignant Sinonasal Tumors

Objective:To explore the value of the predictive model of radiomic nomogram based on magnetic resonance imaging(MRI)in differentiating sinonasal benign and malignant tumors.Materials and methods:A retrospective analysis of 197 cases of sinonasal tumor patients treated by the Affiliated Hospital of Qingdao University from March 2006 to June2019.Among them,113 patients with benign tumors(75 males,38 females,mean age54.7±14.1 years)and 84 patients with malignant tumors(54 males,30 females,mean age55.0±17.4 years).The examinations of all patients were performed on our institution’s GE3.0T MRI and 1.5T MRI,the imaging sequences included axial fast-spin-echo(FSE)T1WI sequence and axial FSE T2WI fat-suppressed(FS)sequence.They were divided into training group(n=138/3.0T MRI)and validation group(n=59/1.5T MRI)according to different MRI.Clinical features(age,gender)and MRI morphologic features(MRI signal matrix,size,margin,myxoid matrix,necrosis matrix,septations,bone involvement,etc.)were retrospectively reviewed for patients in both tumor categories,all data were analyzed using IBM SPSS 20.0,the continuous variables were tested by independent t test orWilcoxon test,and the categorical variables were tested byχ~2test or Fisher’s exact probability test.Firstly,the axial T1WI and T2WI FS DICOM format images wereextracted from the Picture Archiving and Communication Systems(PACS)system of our institution and imported into ITK-SNAP software,the largest layer of the tumor in theMRI images of all patients was manually outlined to serve as the region of interest(ROI)of the images,A.K software was applied to quantitatively extract the radiomic nomogram features at the largest layer of the tumor in axial T1WI and T2WI FS sequences from the ROI;Then,radiomic features were selected by Minimum Redundancy MaximumRelevance(m RMR)and the least absolute shrinkage and selection operator(LASSO)regression model to construct radiomic signatures;Finally,univariate and multivariate logistic regression analysis was used to identify independent risk factors,and combined with the radiomic signature was used to construct the radiomic nomogram,the area under the curve(AUC),sensitivity,specificity and accuracy of the receiver operating curve(ROC)were used to evaluate the differential diagnostic efficacy of five models(clinical model,T1WI based radiomic signature,T2WI FS based radiomic signature,combined radiomic signature and radiomic nomogram).The difference in AUC between any two combinations of the three models(clinical model,combined radiomic signature andradiomic nomogram)was evaluated by Delong test.Results:Univariate and multivariate logistic regression analysis confirmed that lesion margin and bone involvement were independent risk factors for malignant sinonasal tumors.Using the two sequences analyzed with T1WI,T2WI FS and a combination of both,14 radiomic features were finally retained by lasso regression analysis,including 8 radiomic features from T2WI FS and 6 from T1WI.The AUCs of the radiomic nomogram in identifying benign and malignant sinonasal tumors were 0.91 and0.92 in the training and validation groups,respectively;the AUCs of the combined radiomic signature were both 0.88 in the training and validation groups;the AUCs of the T1WI based single radiomic signature were both 0.81 in the training and validation groups;the AUCs of the T2WI FS based single radiomic signature were both 0.86 in the training and validation groups;and the AUCs of the clinical model were 0.71 and 0.83 in the training and validation groups,respectively.There was no significant difference in AUC between the combined radiomic signature and radiomic nomogram(P>0.05),but the radiomic nomogram showed a relatively higher AUC than the combined radiomic signature.There was a significant difference in AUC between the following two models(radiomic nomogram vs clinical model,both P<0.001;combined radiomic signature vs clinical model,P=0.0252 and 0.0035 in the training and validation groups).The radiomic nomogram had better calibration and clinical net benefit.Conclusions:The radiomic nomogram based on MRI plain scan T1WI and T2WI FS sequences had a good predictive effect in differentiating benign and malignant sinonasal tumors preoperatively.