Application Of Multiparametric Magnetic Resonance Imaging-Based Radiomics In The Preoperative Prediction Of Pathological Features And Molecular Subtype-Related Features Of Bladder Cancer

Part 1 Multiparametric magnetic resonance imaging-based radiomics signature for preoperative prediction of muscle-invasive status in bladder cancerBackground: Bladder cancer(BCa)is the 10 th most prevalent malignancy and the most common urological cancer.According to the muscle-invasive status,BCa is clinically divided into nonmuscle-invasive bladder carcinoma(NMIBC)and muscleinvasive bladder carcinoma(MIBC),which is critical in progress prediction and treatment decision making.Currently,the cystoscopic biopsy is commonly used to recognize the muscle-invasive status of BCa.However,this approach is invasive with a risk of misdiagnosis.Thus,setting up a non-invasive and accurate method to preoperatively predict the muscle-invasive status is needed.Radiomics can extract high throughput and quantitative radiomics features that cannot be deciphered by human.These radiomics features may contain information about the biological characteristics of the tumor,so it is possible to construct radiomics signatures for the preoperative evaluation of the biological behavior and heterogeneity on the onset of tumor.Purpose: We aimed to extract radiomics features from the preoperative multiparametric MRI(mpMRI)of BCa patients,and build radiomics signatures to preoperatively predict the muscle-invasive status in BCa.In addition,a nomogram based on the radiomics signature and clinical factor was constructed to improve the performance.Method: One hundred and eighty-five BCa patients from August 2014 to April 2020 in Shanghai Tenth People’s Hospital were retrospectively collected and randomly divided into the training set(n=129)and the validation set(n=56)at a ratio of 7:3.Radiomics features were extracted from the dynamic contrast-enhancement(DCE)and T2 WI sequences of preoperative mpMRI.The intra-and interclass correlation coefficients and the minimum redundancy maximum relevance were used for feature screening.The selected features were introduced to construct three radiomics signatures,including random forest(RF),support vector machines(SVM)and least absolute shrinkage and selection operator(LASSO)in the training set to preoperatively predict the muscle-invasive status.Univariable and multivariable logistic regression were analyzed to construct a nomogram based on the optimal radiomics signature and clinical characteristics.The area under the curve(AUC),accuracy and calibration curve were used to evaluate the performance of the model.Decision curve analysis(DCA)was used to investigate the clinical utility of the model.Results: In total,1,218 radiomics features were quantitatively extracted from the T2 WI and DCE sequences,respectively.The LASSO model had the best capacity for muscle invasive status prediction,with the accuracy of 90.7% and 89.29% in training set and validation set,respectively,and the AUC of 0.934 and 0.906 in training set and validation set,respectively.The results of multivariable logistic regression showed that the LASSO model and the Vesical Imaging-Reporting and Data System(VI-RADS)score were the independent factors for the muscle-invasive status.A nomogram based on the LASSO model and the VI-RADS score demonstrated better discrimination of muscle-invasive status with the AUC of 0.970 and 0.943 in training set and validation set,respectively.DCA and calibration curve further presented the clinical utility and the good performance of the nomogram,respectively.Conclusion: The mpMRI-based radiomics signature may be useful and non-invasive way to preoperatively differentiate the muscle-invasive status in BCa.The proposed nomogram integrating the radiomics signature with the VI-RADS score further improve the differentiation power.Part 2 Multiparametric magnetic resonance imaging-based radiomics signature for preoperative prediction of Ki67 expression in bladder cancerBackground: The Ki67 expression is associated with the muscle-invasive status,tumor grade,N stage,lymphovascular invasion status and prognosis in BCa.For NMIBC patients who were treated with transurethral resection of bladder tumor(TURBT)and Bacillus Calmette-Guérin intravesical immunotherapy,high expression of Ki67 was associated with poorer progression-free survival.Thus,Ki67 expression is not only a useful indicator of prognoses and tumor characteristics,but also may be a reference tool for treatment decision making.Currently,the Ki67 expression can only be detected by immunohistochemistry(IHC)using tumor tissues from either cystoscopic biopsy or radical cystectomy.However,this process is invasive,and the destruction of tumor tissue by surgical instruments may affect the results of IHC.Therefore,setting up a non-invasive and accurate method to predict the Ki67 expression is needed.Radiomics is a non-invasive approach that can extract high throughput and quantitative radiomics features.These radiomics features may contain information about the biological characteristics of the tumor.Purpose: We aimed to extract radiomics features from the preoperative multiparametric MRI and develop radiomics signatures to preoperatively predict the Ki67 expression status in BCa.In addition,the prognostic value of the IHC-based Ki67 expression and the radiomics-based Ki67 expression were further investigated.Methods: We retrospectively collected 179 BCa patients from August 2014 to April2020 in Shanghai Tenth People’s Hospital and randomly divided these patients into the training set(n=125)and the validation set(n=54)at a ratio of 7:3.We extracted radiomics features from the DCE and T2 WI sequences of preoperative mpMRI.In the training set,the intra-and interclass correlation coefficients and the minimum redundancy maximum relevance were used for feature screening,and the synthetic minority over-sampling technique(SMOTE)was used to balance the minority group(low Ki67 expression group).The SVM and LASSO models were built in the training set,and the SMOTE-SVM and SMOYE-LASSO model were built in the SMOTEtraining set.These four models were validated in the validation set to evaluate the performance of these models in preoperatively predicting the Ki67 expression status.The AUC,accuracy and calibration curve were used to evaluate the performance of the model.DCA was used to investigate the clinical utility of the model.KaplanMeier analysis was used to investigate the prognostic value of the IHC-based Ki67 expression and the radiomics-based Ki67 expression.Results: In total,1,218 radiomics features were quantitatively extracted from the T2 WI and DCE sequences,respectively.The SMOTE-LASSO model based on nine radiomics features achieved the best performance in predicting the Ki67 expression status,with the AUC of 0.859 and 0.819 in training and validation sets,respectively,and the accuracy of 80.3% and 81.5% in training and validation sets,respectively.DCA and calibration curve further presented the clinical utility and the good performance of the SMOTE-LASSO model,respectively.IHC-based and radiomicsbased Ki67 expression had prognostic value in both training and validation sets.BCa patients with IHC-based or radiomics-based high Ki67 expression were significantly associated with poor disease-free survival(P value all <0.05).Conclusion:The SMOTE-LASSO model based on the preoperative mpMRI radiomics features could predict the expression of Ki67 and was related to the prognoses of BCa patients,thereby may aid in clinical decision-making.Part 3 Multiparametric magnetic resonance imaging-based radiomics signature for preoperative prediction of molecular subtype-related features in bladder cancerBackground: Currently,various published molecular classifications have been developed for BCa.According to these molecular classifications,BCa could be roughly divided into basal and luminal subtypes.The molecular subtypes of BCa can be detected by IHC or RNA-sequence,which is invasive.In addition,the qualitative molecular classifications may not accurate and comprehensive enough to describe the molecular biological characteristics of the BCa.Thus,it is necessary to develop a quantitative biomarker to assess the molecular subtype-related features in BCa,and develop a new approach to preoperatively predict the molecular subtype-related features of BCa patients.Purpose: We aimed to develop a quantitative biomarker(named as basal-luminal score)to assess the molecular subtype-related features in BCa,and investigate its potential biological mechanisms and association with the molecular classifications,prognosis and drug susceptibility.In addition,we extracted radiomics features from the preoperative mpMRI of BCa patients and developed mpMRI-based radiomics signatures to preoperatively predict the basal-luminal score.Methods: The clinical data and RNA-sequence data were downloaded from the TCGA,GEO and IMvigor210 datasets.‘BLCAsubtyping’ and ‘consensus MIBC’ R packages were used to obtain the results of seven published molecular classifications.‘p RRophetic’ R package was used to predict the half maximal inhibitory concentration(IC50)of common chemotherapy drugs in BCa.Gene Set Variation Analysis(GSVA)was performed to evaluate the basal-luminal score based on RNA expression data of gene sets for basal and luminal features.The associations of the basal-luminal score with seven molecular classifications,prognosis and drug susceptibility were investigated.Kaplan-Meier analysis,univariate/multivariate Cox regression analysis and meta-analysis were used to investigate the prognostic value of the basal-luminal score.Gene Set Enrichment Analysis(GSEA)were used to investigate the biological mechanisms of the basal-luminal score.A total of 111 BCa patients with RNA-sequence data of tumor tissue and preoperative mpMRI in Shanghai Tenth People’s Hospital between November 2019 and July 2021 were obtained in this study.IHC and western blot based on tumor tissues were performed to validate the biological mechanisms of the basal-luminal score and the association between the basal-luminal score and the molecular classifications.These BCa patients in our center were randomly divided into the training set(n=77)and the validation set(n=34)at a ratio of 7:3.Radiomics features were extracted from the DCE and T2 WI sequences in each patient.To predict the basal-luminal score,three radiomics signatures,including LASSO,SVM and RF models,were constructed in the training set and validated in the validation set.The AUC and accuracy were used to evaluate the performance of the model.Results: The basal-luminal score was highly associated with seven molecular classifications.Patients with high basal-luminal score were mainly allocated into basal-related subtypes of molecular classifications.The results of IHC in our center revealed that patients with high basal-luminal score had high expression of basalrelated biomarkers,including CK5/6 and CK14.On the contrary,patients with low basal-luminal score had high expression of luminal-related biomarkers,including GATA3,CK20 and FOXA1.The results of Kaplan-Meier analysis,univariate/multivariate Cox regression analysis and meta-analysis showed that basalluminal score was an independent risk biomarker,and BCa patients with high basalluminal score had poor survival outcomes.The results of drug susceptibility suggested that BCa patients with high basal-luminal score had significantly lower IC50 value of four common chemotherapy drugs in BCa,revealing that these patients with high basal-luminal score may be more sensitive to chemotherapy.GSEA showed that several oncogenic pathways,including pathways in cancer,WNT signaling pathway,transforming growth factor-β(TGF-β)signaling pathway,cell cycle and bladder cancer,were mainly enriched in high basal-luminal group.The results of IHC(Cyclin D1 and β-cadherin)and western blot(EGFR,β-cadherin,p-SMAD2/3,Cyclin D1 and Bcl2)in our center further validated the results of GSEA.A total of3,562 radiomics features were extracted from the DCE and T2 WI sequences(1,781 features per sequence)in each patient.After features screening using the intra-and interclass correlation coefficients and the minimum redundancy maximum relevance,a RF model based on nine radiomics features achieved the optimal performance in predicting the basal-luminal score,with the accuracy of 81.8% and 85.3% in training and validation sets,respectively,and the AUC of 0.894 and 0.823 in training and validation sets,respectively.Conclusion: The basal-luminal score was associated with seven molecular classifications and could quantitatively reflect the molecular subtypes of BCa.High basal-luminal score was related to basal features,and low basal-luminal score was related to luminal features.The basal-luminal score had the potential to predict the survival outcomes and chemosensitivity.Several oncogenic pathways were highly enriched in patients with high basal-luminal score,which partially explained the poor survival outcomes of these patients.The mpMRI-based radiomics signature may be useful for preoperatively predicting the basal and luminal features in BCa,thereby aided in clinical decision-making.