Application Of Radiomics And Texture Analysis In Predicting Stages And Phenotypes Of Patients With Rectal Cancer

Part 1:The application of functional MRI and T2-based texture analysis in differentiating T1-2 and T3a rectal cancerPurpose:This study aimed to investigate whether the texture features could differentiate rectal cancer from pathological stage of T1-2（pT1-2）to T3a（pT3a）.Methods:82 patients diagnosed as rectal adenocarcinoma of pT1-2 and pT3a stages were enrolled into the study,and were grouped as pT1-2 and pT3a.Thirty patients were at the stage of pT1-2,while fifty-two patients were at the stage of pT3a.All patients received regular T2-weighted imaging and functional magnetic resonance imaging（fMRI）examination,which included apparent diffusion coefficient（ADC）sequence,dynamic contrast enhancement（DCE）MRI and intravoxel incoherent motion diffusion weighted imaging（IVIM）.Acquired parameters are:ADC values from ADC sequence,transfer constant（Ktrans）,reflux constant（Kep）,extravascular extracellular fractional volume（Ve）from DCE sequences,and perfusion fraction（f）,true diffusion coefficient（D）,pseudodiffusion coefficient（D*）,from IVIM sequence.The texture features were obtained by delineating the regions of interest（ROI）on the largest cross-sectional aera from T2 image.The accuracy in differentiating pT1-2 and pT3 a rectal cancer before surgery by doctors using magnetic resonance imaging（MRI）and the area under the curve（AUC）of doctors’ diagnosis were calculated.The differences of clinicopathological variables,functional MRI parameters and texture features were compared between the groups using univariate analysis（P<0.05）.Correlation of texture features,and functional parameters with stage（grouped as pT1-2 and pT3a）were calculated by using r（Spearman’s rank correlation coefficient）.Finally,receiver operating characteristic（ROC）curves of selected features were generated to distinguish pT1-2 and pT3a tumors of rectal cancer.DeLong test was used to compare the receiver operation characteristic curves of texture features and functional MRI parameters.Results:The preoperative accuracy of doctors in differentiating pT1-2 and pT3a rectal cancer was shown to be 74.39%,with the aera under the curve of 0.563.The respective values of Kep,Ve,and ADC showed significant differences between the groups（PKep=0.013,PVe=0.022,PADC=0.004）.The Kep and ADC values showed negative correlation with stage（rKep=-0.277,rADC=-0.318）.Ve showed positive correlation with stage（rVe=0.255）.Twenty five texture features from T2 images showed significant differences between groups,and S（0,2）SumOfSqs and WavEnLH_s2 showed best performance in differentiating pT1-2 from pT3a rectal cancer,with AUC of 0.721 and 0.699 respectively.S（0,2）SumOfSqs and WavEnLH_s2 demonstrated negative correlation with tumor stage（r=-0.369,r=-0.332）.The area under the curve of ADC,Kep,and Ve,were,0.690,0.666,0.653 respectively.The combined AUC,sensitivity and specificity of the two texture features,Kep,Ve and ADC were 0.833,88.46%,and 73.33%.Also,DeLong test and ROC curves showed that the performance of joint prediction was better than doctors’ preoperative diagnosis,single texture parameter,and single functional MRI parameter.And there was a statistical difference（P<0.05）between them.What’s more,the area under the curve,sensitivity and specificity didn’t get improved after incorporating the doctor’s diagnosis results into the above combination.Conclusion:Texture analysis is able to differentiate the stages of rectal cancer.And the performance of texture features with Ve,Kep and ADC in differentiating pT1-2 from pT3a tumors in rectal cancer patients is superior to the single method.The application of this combination could be beneficial to follow-up accurate and individualized treatment strategies for patients with rectal cancer.Part 2:The application of T2-based radiomics identifying HER-2 positive patients of rectal cancerPurpose:To find out the ability of radiomics in predicting human epidermal growth factor receptor-2（HER-2）positive patients with rectal cancer.Methods:130 patients diagnosed as rectal adenocarcinoma who have T2 and apparent diffusion coefficient images from magnetic resonance imaging were enrolled into the study.Patients were divided into HER-2 positive and negative groups,according to immunohistochemistry（IHC）and Fluorescence in situ hybridization（FISH）results,and were then divided into a training cohort and a validation cohort at a ratio of 7:3.91 patients were enrolled into the training cohort,with 26 HER-2 positive patients,and 65 HER-2 negative patients.39 patients were divided into the validation cohort,with 12 HER-2 positive patients and 27 negative patients.Radiomics features were acquired by delineating regions of interest onT2 images,and the method of Least absolute shrinkage and selection operator（LASSO）was applied to filter the radiomics features and to acquire radiomics signature.Uni-variate analysis was applied to acquire clinicopathological features with significance（P<0.05）.Then,logistic regression analysis incorporating radiomics signature and clinicopathological features was applied to build the clinical-radiomics model and to find out the independent risk factors.Receiver operating characteristic curves were generated to see predictive performance of the clinical-radiomics model and the radiomics signature.DeLong test was applied to find out the differences of ROC curves between clinical-radiomics model and radiomics signature.Calibration curves and Hosmer-Lemeshow test were used to show the fitness of the model.Decision curve analysis（DCA）was used to assess the clinical usefulness of the model.The model built in the training cohort was then validated in the validation cohort.Results:6 radiomics features were selected to build the radiomics signature.Differentiation grade of the rectal cancer showed statistical significance between groups（P=0.037）.Radiomics signature is the independent risk factor of the HER-2 positive expression.The area under the curve in the training cohort of the radiomics signature and the clinical-radiomics predicting model were 0.893 and 0.899,with the sensitivity of 76.92%and 76.92%,the specificity of 90.77%and 93.85%.In the validation cohort,the AUC,sensitivity and specificity of the radiomics signature and the clinical-radiomics predicting model were 0.713、75.00%,74.07%and 0.735,66.67%,88.89%.The DeLong test showed no statistical difference between the radiomics signature and the clinical-radiomics predicting model in both training and validation cohort.However,the specificity of the clinical-radiomics predicting model was improved compared with the radiomics signature in both training and validation cohort.The Hosmer-Lemeshow test showed good fit of the model in the training cohort（P=0.677）,and the moderate fitness in the validation cohort（P=0.317）.DCA showed that the clinical-radiomics predicting model was useful for the clinical practice.Conclusion:The radiomics signature and the clinical-radiomics predicting model are able to predict the positive expression status of HER-2 in rectal cancer patients,with the latter having improved specificity,enhancing our confidence in diagnosis.The radiomics signature and the clinical-radiomics predicting model both could provide quantitative information and guide accurate and individualized treatments during clinical practice.