| Part Ⅰ: Automatic segmentation of bladder cancer on DWI: application of deep-learning convolutional neural networkObjective: To develop a V-Net based deep-learning convolutional neural network(CNN)to automatically segment bldder cancer on diffusion-weighted imaging(DWI).Materials and Methods : This retrospective study included 200 pathologically confirmed bladder cancer patients(164 males,62.1±10.6 yrs [37-79])who underwent DWI before transurethral resection of bladder tumor between July 2015 and July 2018 at our institute.Patients were randomly divided into training set(n=150)and validation set(n=50).One radiologist manually segmented the entire tumor area on DWI(b = 1000 s/mm2)to yield volume of interest,namely ground truth.We constructed improved V-Net models by adding skipping connection architecture on the original V-Net model: one original V-Net architecture added skipping connection architecture to construct single-stage improved V-Net model;two original V-Net architectures were connected in series and then skipping connection architecture was added to construct multi-stage improved V-Net model.In the training set,original V-Net model and two improved V-Net models were trained from scratch by employing the Adam optimizer with standard β equal to 0.9 and 0.99.The learning rate was initially set to 0.05,and then decreased to 0.97 fold each time encountering 20 epochs with a total of 6000 epochs.All models were implemented in Py Torch v1.2,and dice similarity coefficient(DSC)was used to evaluate their performance for automatic image segmentation in the validation set.Results: The DSC of the single-stage improved V-Net model for automatic image segmentation was better(0.90447±0.00411 vs 0.87806±0.01139,p<0.001,Wilcoxon signed-rank test)than that of the original V-Net model.The DSC of the multi-stage improved V-Net model was better than both the original V-Net model(0.95217±0.00122 vs 0.90447±0.00411,p<0.001)and the single-stage improved V-Net model(0.95217±0.00122 vs 0.87806±0.01139,p<0.001).Conclusions:Compared with original V-Net model,our improved V-Net models can yield more precise segmentation of bladder tumor on DWI in bladder cancer patients.The whole segmentation process is fully-automatic and yields results similar to the ground truth,demonstrating the ability of CNN for automatic segmentation of bladder cancer DWI images.Part Ⅱ: Combining DWI based radiomics features with transurethral resection of bldder tumor promotes the differentiation between muscle-invasive bladder cancer and non-muscle-invasive bladder cancerObjective:With the result of transurethral resection of bladder tumor (TURBT),this study aimed to develop and validate a more sensitive radiomics model from diffusion-weighted imaging(DWI)for discriminating muscle-invasive bladder cancer(MIBC)from non-muscle-invasive bladder cancer(NMIBC).Materials and Methods : This retrospective study included 218 pathologically confirmed bladder cancer patients(169 males,66.1 yrs [37-93];141 MIBC)who underwent DWI before TURBT between July 2015 and December 2018 at our institute.All patients underwent both TURBT and radical cystectomy,and had standard pathologic staging results.Patients were randomly divided into training set(n=131;104 males,65.8 yrs [38-86];86 MIBC)and validation set(n=87;65 males,66.5 yrs [37-93];55 MIBC).Volume of interest(VOI)was obtained by manually segmenting tumor area on DWI(b = 1000 s/mm2),and copied to corresponding apparent diffusion coefficient map.Then extracted the radiomics features within the VOI,and built a classification model(RF1)using random forest algorithm for discriminating MIBC from NMIBC.Similarly,we also extracted the radiomics features within the VOI automatically segmented by V-Net based deep-learning convolutional neural network,and built another classification model(RF2).Combination models based on TURBT results and classification models were also built.All the models were trained in the training set and tested in the validation set.In all tests,MIBC was regarded as the positive result.Discrimination performances were evaluated with the area under the receiver operating characteristic(ROC)curve(AUC),accuracy,sensitivity,specificity,F1 and F2 scores.Qualitative MRI evaluation based on morphology was performed for comparison.Delong’s test was used for comparing AUC,and Mc Nemar’s test for comparing ACC,SEN and SPE between the two models.Results:No significant difference was found in AUC(0.907 vs 0.901,p=0.540,Delong’s test),ACC(0.839 vs 0.812,p=0.580,Mc Nemar’s test),SEN(0.873 vs 0.845,p=1.000)and SPE(0.781 vs 0.740,p=1.000)between RF1 and RF2 for discriminating MIBC from NMIBC.RF1 was more sensitive than TURBT(0.873 vs 0.655,p=0.019),and MRI(0.873 vs 0.764,p=0.181),but the difference did not reach statistical significance.When combining the RF1 with TURBT,the sensitivity increased to 0.964,significantly higher than TURBT(0.655,p<0.001),MRI(0.764,p=0.006),and the combination of TURBT and MRI(0.836,p=0.046).Notably,the combination model(RF1 and TURBT)had the highest accuracy of 0.897 and F2 score of 0.946 for discriminating MIBCConclusion : Combining DWI radiomics features with TURBT could improve the sensitivity and accuracy in discriminating MIBC from NMIBC for clinical practice.And the application value of computer-aided diagnosis technology in clinical diagnosis process was preliminarily explored in this study.Multi-center,prospective studies are needed to confirm our results.Part Ⅲ: DWI based radiomics nomogram predicting progression-free survival in patients with muscle-invasive bladder cancerObjective:We hypothesized that diffusion-weighted imaging(DWI)based radiomics features have potential correlation with the progression risk of muscle-invasive bladder cancer(MIBC)patients.Based on this hypothesis,we constructed a radiomics nomogram combining radiomics signature and clinicopathologic characteristics to predict the progression free survival(PFS)of patients,and to evaluate the incremental value of radiomics signature over traditional clinicopathologic risk factors.Materials and Methods: This retrospective study included 210 pathologically confirmed bladder cancer patients(162 males,66.0±9.6 yrs [37-84])who underwent DWI before transurethral resection of bladder tumor(TURBT)between July 2014 and December 2018 at our institute.All the patients were diagnosed as MIBC by the combination of TURBT and MRI.Patients were randomly divided into training set(n=105)and validation set(n=105).Volume of interest(VOI)was automatically segmented by V-Net based deep-learning convolutional neural network on DWI(b = 1000 s/mm2),and copied to corresponding apparent diffusion coefficient map.Then the radiomics features within VOI were extracted.In the training set,predictability analysis and LASSO regression model were used to reduce the dimension of the data,and to select the best subset of radiomics features for constructing radiomics signature.In the training set and the validation set,Kaplan-Meier survival analysis was performed to evaluate the potential correlation between radiomics signature and PFS.Clinicopathologic characteristics including gender,age,neoadjuvant chemotherapy status,radical cystectomy status,Ki-67,clinical lymph node metastasis,clinical stage and concomitant CIS were combined with radiomics signature to build multivariate Cox proportional risk regression model for predicting PFS to evaluate whether radiomics signature could be used as independent predictors of PFS.Clinicopathologic nomogram and radiomics nomogram with PFS as the end point were constructed respectively.The independent variables of the clinicopathologic nomogram only included clinicopathologic characteristics,while the radiomics nomogram incorporated with clinicopathologic characteristics and radiomics signature as independent variables.The performance of the radiomics nomogram and clinicopathologic nomogram was compared,and the incremental prognostic value of the radiomics signature over traditional clinicopathologic risk factors was assessed in terms of discrimination,calibration,and clinical usefulness.Results: There was a significant correlation between radiomics signature and PFS in both the training set(hazard ratio [HR]=4.55;95% confidence interval [CI]: 2.27,9.11;p=0.00017)and the validation set(HR=1.85;95% CI: 0.96,3.56;p=0.0073).In the validation set,multivariate Cox proportional risk regression analysis showed that radiomics signature was an independent predictor of PFS(HR=2.57;95% CI: 1.53,4.33;p=0.0004).Compared with clinicopathologic nomogram,the predictive results of radiomics nomogram were better consistent with the actually observed results,and the predictive accuracy is higher(net reclassification improvement =0.226,95% CI: 0.016,0.415,p=0.038),suggesting that radiomics signature can improve the efficiency of prediction model.Decision curve analysis showed that the radiomics nomogram achieved a higher overall net benefit than the clinicopathologic nomogram within most range of threshold probabilities.Conclusion: The DWI based radiomics signature constructed in this study was an independent predictor of PFS in MIBC patients.Compared with clinicopathologic nomogram,the predictive results of radiomics nomogram were better consistent with the actually observed results,and the predictive accuracy is higher.The progression free survival probability of each patient given by the nomogram can provide reference for making individualized treatment strategy.Multi-center,prospective studies are needed to confirm our results.. |
PDF Full Text Request