Liposarcoma:Predictive Histopathological Subtypes On Radiomics Of MRI

Objective:To explore the feasibility and value of predicting histopathological subtypes of liposarcoma by Radiomics model based on T₁WI and FS-T₂WI.Materials and Methods:1.General informationPatients who underwent MRI examination and pathologically diagnosed at the Second Hospital of Da Lian Medical University between June 2013 to January 2020were recruited in this retrospective study.76 patients were finally collected,including41 males and 35 females,aged 21-85.The average age was 58±13.7 years.73 cases were underwent fat suppressed?FS-T₂WI?,including atypical lipomatous tumor/well-differentiated 20,myxoid 27,dedifferentiated 14,and pleomorphic 12;71 cases were underwent T₁WI,including atypical lipomatous tumor/well-differentiated 18,myxoid 28,dedifferentiated 11,and pleomorphic 14.2.Image obtaining and processing2.1 MRI equipment and methodMagnetic resonance scanner included GE Discovery MR 750W 3.0T?Siemens Verio 3.0 T?GE HDxt 1.5T,using Keen joint coil,large flexible coil,head-neck joint coil or 16-channle phased array coil.The sequences included T₁WI,FS-T₂WI,and DWI and contrast injection in some cases.2.2 The analysis of conventional imagingConventional MRI of liposarcoma was analyzed by two radiologists?radiologist A and B with 3 and 8 years of musculoskeletal diagnosis experience,respectively?.Analysis characteristics of tumors according to the size of the tumor?longest diameter of the tumor?,the location of the tumor,the boundary?clear/unclear?,the characteristics of the tumor plain signal,the DWI signal,the T₂WI heterogeneity,and the tumor enhancement.2.3 Analysis of radiomics?1?Delineated the volume of interest?VOI?of tumorThe MRI of DICOM format is uploaded to the Radiomics cloud platform ?http://mics.radcloud.com?of Huiying Medical Technology?Beijing?Co.,Ltd.The VOI of the tumors was boundary delineated layer by layer on T₁WI and FS-T₂WI respectively and subjected to three-dimensional volume reconstruction.The VOI was included cystic changes,necrosis and bleeding areas.?2?Features extractionThe VOI of the tumors on T₁WI and FS-T₂WI was calculated and features included feature class and filter-based class.Feature class was divided into first order statistics,shape-and size-based features,and texture features.Filter-based features include wavelet transformation and Laplacian Gaussian filtering.?3?Dimensionality reductionThe features were filtered by Variance Threshold,Select KBest and Principal Component Analysis in different sequences.?4?Establish classification models80%cases of each subtype were trained by K-nearest neighbor?KNN?and support vector machine?SVM?on T₁WI and FS-T₂WI respectively,and 20%cases were tested by the two classifiers.Two classification models were established finally.?5?Statistical analysisPlot the ROC curve and compare the area under the curve?AUC??sensitivity?specificity to evaluate the diagnostic performance of each model.The closer the AUC is to 1,the better the diagnostic performance of the model.F1-score was used to evaluate the stability of models.The closer the F1-score is to 1,the better stability the model has.Results:1.MRI characteristics of liposarcoma subtypesTypical liposarcoma subtypes have their own characteristics,atypical lipomatous tumor/well-differentiated liposarcoma mostly located in the thigh,and it appears as an enveloped and separated mature fat mass.Myxoid liposarcoma is characterized as high mucus-containing,high T₂WI signal intensity and low T₁WI signal intensity,and a little mature fat can be seen in some cases.Dedifferentiated liposarcoma often located in the retroperitoneum.The tumor is composed of mature fat and highly heterogeneous sarcomas,and the two parts are interrupted suddenly.Pleomorphic liposarcoma may show heterogeneous soft tissue sarcomas on imaging,often accompanied by hemorrhage and necrosis,with little or no adipose.2.The results of radiomics2.1 The results of feature extraction and dimensionality reduction1412 features were extracted from the VOI of tumors on T₁WI and FS-T₂WI respectively.After Variance Threshold,the features were reduced to 898 and 924 respectively,after Select KBest,the features were reduced to 120 and 258 respectively.Finally,analysis and reduce features by Principal Component Analysis.2.2 Results of classifier models?1?Results of KNN classifier based on T₁WI:AUC of atypical lipomatous tumor/well-differentiated is 0.83,F1-score is 0.60;AUC of myxoid is 0.87,F1-score is 0.74;AUC of dedifferentiated is 0.78,F1-score is0.55;AUC of pleomorphic is 0.76,F1-score is 0.57;?2?Results of SVM classifier based on T₁WI:AUC of atypical lipomatous tumor/well-differentiated is 0.93,F1-score is 0.73;AUC of myxoid is 0.90,F1-score is 0.86;AUC of dedifferentiated is 0.83,F1-score is0.67;AUC of pleomorphic is 0.76,F1-score is 0.72;?3?Results of KNN classifier based on FS-T₂WI:AUC of atypical lipomatous tumor/well-differentiated is 0.96,F1-score is 0.72;AUC of myxoid is 0.88,F1-score is 0.80;AUC of dedifferentiated is 0.77,F1-score is0.57;AUC of pleomorphic is 0.84,F1-score is 0.67;?4?Results of SVM classifier based on FS-T₂WI:AUC of atypical lipomatous tumor/well-differentiated is 0.90,F1-score is 0.76;AUC of myxoid is 0.96,F1-score is 0.86;AUC of dedifferentiated is 0.90,F1-score is0.71;AUC of pleomorphic is 0.88,F1-score is 0.82;Conclusion:The following conclusions can be drawn by analyzing the T₁WI and FS-T₂WI radiomics of 76 cases of different histopathological subtypes liposarcoma:1.This study is the first application of radiomics to analyze the difference of liposarcoma histopathological subtypes.Radiomics base on MRI provides a new assistant diagnostic method for predicting liposarcoma subtypes.2.T₁WI-based KNN classifiers,T₁WI-based SVM classifiers,FS-T₂WI-based KNN classifiers,FS-T₂WI-based SVM classifiers can identify the histopathological subtypes of liposarcoma.Among them,the FS-T₂WI-based SVM classifier has the best diagnostic performance,AUC of atypical lipomatous tumor/well-differentiated is 0.90;AUC of myxoid is 0.96;AUC of dedifferentiated is 0.90;AUC of pleomorphic is 0.88.