Research On The Technology Of Preoperative Prediction Of Muscular Invasiveness And Stage Of Bladder Cancer Via 3D Features

In 2015,CA cancer statistics reported that urinary bladder cancer(BC)has become the fourth most frequently occurring tumor and the ninth most common cause of cancer-related deaths among males worldwide.The next year,their annual cancer statistics revealed that BC has become the seventh most frequently occurring cancer and the twelfth most common cause of cancer-related deaths among males in China.According to the National Comprehensive Cancer Network(NCCN)guideline for bladder cancer,patients with different muscle-invasive depth(T stage)and histo-pathological grade may be treated differently with different follow-up strategies.Therefore,accurate prediction of muscular invasiveness of BC preoperatively is of great clinical importance for therapeutic and follow-up decisions.Currently,optical cystoscopy(OCy)and histo-pathological examination of transurethral resection(TUR)biopsies are standard experience for BC detection and diagnosis.Literatures indicate that this diagnosis strategy might cause some errors that a significant portion of muscle-invasive bladder carcinomas(MIBC,T stage ? T2)have been understaged.Repeated examinations with OCy and TUR biopsies may reduce the error,but the effect is limited due to its invasive,uncomfortable,time-consuming and costly procedures.Therefore,a convenient and non-invasive technology to discriminate between BC and normal wall tissues,followed by preoperative differentiation of MIBC from non-muscle-invasive bladder carcinoma(NMIBC,T stage ? T1),and quantitatively assessing the invasion depth of BC via medical images,would be very beneficial for treatment planning and follow-up management.Compared with traditional OCy and TUR procedures,image-based computer assisted diagnosis(CADx)on bladder abnormalities,which is non-invasive,safe,and more comfortable,has revealed its great potential in differentiation of cancerous and nornal tissues,and of benign and malignant cancerous tissues.Recently,extensive studies have reported that three-dimensional(3D)textural features are widely applied in characterizing tumor heterogeneity,showing more effective in differentitiation of benign from malignant than the two-dimensional(2D)features.Currently,3D texture features have been extensively used in the prediction and diagnosis of colorectal cancer,breast tumors and lung nodes.As for bladder tumor,(1)whether 3D texture features outperform 2D versions in distinguishing bladder tumors from normal wall tissues,(2)what kinds of texture features could better reflect the heterogeneous distribution of BC,(3)whether the features characterize heterogeneity of BC exist significant differences between muscle-invasive(stage ? T2)and non-muscle-invasive(stage ? T1)BC,(4)how to apply these features with significant differences for cancer invasiveness differentiation,and(5)how to futher realize the quantitative evaluation of tumor staging,these questions are currently open and hot issues desirably to be solved in CADx of BC.To address the questions above,in this study,we mainly work on the following parts:Part I Differentiation of cancerous from non-cancerous tissues via MRI texture featuresBased on the T2-weighted MR images(T2WI)and their high-order derivative maps,we analyzed the different performance of using 2D and 3D features in tumor differentiation.Main research works are as follows:(1)Image data acquisition and VOI delineationA total of 62 patients were enrolled in this retrospective study,among them 58 patients were male and 4 were women.All of them were previously underwent T2-weighted MR imaging(T2WI)and pathologically confirmed BC.Sixty-two cancerous and 62 wall volumes of interests(VOI)were manually extracted from T2 WI datasets of the patients by radiologists.(2)High-order derivative map generation and feature extractionTo better reflect heterogeneous distribution of tumor tissues,3D high-order derivative maps(the gradient and curvature maps)were calculated from each VOI.Then 3D Haralick features based on intensity and high-order derivative maps,namely the gray level co-occurrence matrices(GLCM)features,the gray level gradient co-occurrence matrices(GLGCM)features and the gray level curvature co-occurrence matrices(GLCCM)features,together with Tamura features based on intensity maps were extracted from each VOI.For comparison,the corresponding 2D features were also extracted slice by slice from each VOI and then averaged as the final value of that feature.(3)Feature selection and classification verificationStatistical analyses were performed to select the features that have great power to reflect the differences between bladder carcinomas and wall tissues.In order to decrease the feature redundancy for a step further,support vector machine classifier based recursive feature elimination(RFE-SVM)algorithm was proposed to first select a feature subset with optimal compactness and predictive power,and then verify its performance in the differentiation.The results show that: a total of 58 texture features were derived,and 37 features showed significant inter-class differences(P ? 0.01).With 29 optimal features selected by RFE-SVM,the classification results namely the sensitivity,specificity,accuracy and area under the curve(AUC)of the receiver operating characteristics(ROC)were 0.9032,0.8548,0.8790 and 0.9045,respectively.Part II The differentiation of MIBC from NMIBC via 3D radiomic features extracted from conventional MRI and its high-order derivative mapsOur results in Part I indicate that 3D texture features derived from T2 WI and high-order derivative maps can better reflect heterogeneous distribution of cancerous tissues.Based on this observation,in Part II,we further investigated whether there are significant radiomic feature differences between MIBC and NMIBC,and the 3D raidomic features extracted from T2 WI and its high-order derivative maps are capable of reflecting muscular invasiveness of BC.Detailed methods are as follows:(1)MR data collection and VOI delineationA total of 68 patients were enrolled in this retrospective study.All of them were previously underwent T2 WI and pathologically confirmed BC.A total of 118 cancerous VOI were manually segmented from patients' T2 WI by radiologists,including 84 MIBC and 34 NMIBC.The imbalance rate for samples of these two classes was 2.4706,which meant these two classes were largely imbalanced.(2)High-order derivative map generation and radiomic feature extractionTo better reflect heterogeneous distribution differences between MIBC and NMIBC,the radiomic features quantifying tumor signal intensity and textures were extracted from each VOI and its high-order derivative maps,respectively,to characterize heterogeneity of tumor tissues,including signal intensity histogram based features and CM based Haralick features.(3)Feature selectionStatistical analysis was used to build radiomic signatures with significant inter-group differences of NMIBC and MIBC.REF-SVM was performed to further obtain an optimal subset with the best compactness and predictive power.(4)Sample balanceThe synthetic minority oversampling technique(SMOTE)was proposed to balance the imbalanced sample size,then further select the most predictive and compact signature subset to verify its differentiation capability.(5)Classification validationThe SVM classifier was used to verify the overall performance of these features with ten-fold cross-validation.The results show that: from each tumor VOI,a total of 63 radiomic features were derived and 30 of them showed significant inter-group differences(P ? 0.01).By using the SVM-based feature selection algorithm with rebalanced samples,an optimal subset including 13 radiomic signatures was determined.The area under receiver operating characteristic curve and Youden index were improved to 0.8610 and 0.7192,respectively.Part III The differentiation of MIBC from NMIBC via 3D radiomic features extracted from multi-modal MRIThe final results in Part II suggest that 3D radiomic signatures derived from T2 WI and its high-order derivative maps could reflect the heterogeneous differences between MIBC and NMIBC,further predicting the muscle invasiveness of bladder cancer.Considering that functional MRI has been widely used for imaging aided BC invasiveness diagnosis clinically,in Part III,we focused on radiomic features extracted from multi-modal MRI for BC invasiveness differentiation.(1)Subject selectionIn this retrospective study,we selected 53 patients who were underwent both T2 WI and diffusion-weighted MR images(DWI)scanning previously,and 53 tumors were obtained including 19 NMIBC and 34 MIBC.(2)ROI delineationConsidering the blend region of tumor & wall(BRT&W)tissues might have mixed tissue patterns,we further hypothesize that the features derived from the outer region of tumors and the blend region might have different capability of differentiating MIBC from NMIBC.Therefore,two types of ROIs including ROIs of convex region of tumors(CRT)and BRT&W tissues were delineated on patient's T2 WI,DWI and ADC maps.(3)Radiomic feature extractionThe radiomic features quantifying signal intensity and texture were extracted from each ROI to characterize heterogeneity of target tissues,including histogram features,CM based Haralick features,run length matrices(RLM)based features and signal intensity features like DWI-mean and ADC-mean.(4)Feature analysis and classificationStatistical analysis was performed to build radiomic signatures with significant inter-group differences of NMIBC and MIBC.Then the synthetic minority oversampling technique(SMOTE)and RFE-SVM strategy was performed to first balance the sample size,and then select the most predictive and compact signature subset to evaluate its prediction performance.The results show that: a total of 84 radiomic signatures in the ROIs of blend region were selected,whereas only 9 signatures in the ROIs of outer region were enrolled,suggesting that features derived from the blend region of tumor & wall tissues have more potential for cancer invasiveness differentiation.With optimal signature subset selected,the accuracy,Youden index and area under the receiver operating characteristic curve were 0.811,0.567 and 0.834,respectively.With sample balanced and augmented to 34/34,these metrics were advanced to 0.838,0.676 and 0.910,respectively.Part IV Accurate BC segmentation and invasion depth calculationThe previous three parts mainly focused on the problems about how to differentiate cancerous tissues from normal wall tissues,and how to discriminate between MIBC from NMIBC preoperatively via image-based CADx technology.On this basis,how to quantitatively calculate the invasion depth of BC with MR images,and propose the corresponding staging criteria,to realize the accurate BC segmentation,invasion depth calculation and quantitatively staging diagnosis,are the main scientific problems to be solved in this part.Detailed methods are as follows:(1)Subject selectionPreliminarily,we selected 25 patients' T2 WI datasets from all previous T2 WI datasets obtained.These patients were confirmed with BC.Among them,10 patients were staged as ? T1,10 were staged as T2,five were T3 or higher.(2)Accurate BC segmentationFirst,Adaptive Shape Priori Constrained Level-Sets(ASPCLS)was used to segment bladder wall and tumor region from background tissues and inner urine region,followed by manual corrections by radiologists.After that,Priori-Guided Continuous Max-Flow(PG-CMF)was proposed to accurately segment the tumors from their surrounding wall tissues efficiently.(3)Invasion depth calculation and stagingWith the cancerous region accurately segmented,we proposed the concept of relative invasion depth(DRI).To calculate this metric,the minimal bladder wall thickness(BWT)of the tumor region is divided by the average BWT out of the tumor region.Then 1 minus the result is the final maximal DRI,called DRI for short.According to the calculation results of DRI in 25 datasets and the responding histo-pathological diagnosis,we preliminarily proposed quantitatively staging scheme.The results show that: the average DSC was 88.73%,suggesting that PG-CMF can be used for accurate and efficient segmentation of BC.Main conclusions and noveltiesBased on the results of the four parts,we can obtain these conclusions:(1)This paper improves the 3D textural feature models,and first jointly adopts T2 WI and its high-order derivative maps for the differentiation of bladder cancerous tissues from normal wall tissues,creatively proposing construction method for co-occurrence matrix based on 3D intensity-curvature image pair.The results show that 3D texture features derived from T2 WI and high-order derivative maps can better reflect heterogeneous distribution of cancerous tissues,thus have more power in the differentiation of bladder cancerous tissues from normal wall tissues than the 2D ones.(2)This paper proposes the conventional MRI based radiomic strategy for muscular invasiveness differentiation,systematically evaluating 3D radiomic features extracted from T2 WI and its high-order derivative maps for the differentiation task.Results indicate that radiomic features derived from T2 WI and its high-order derivative maps could reflect the heterogeneous differences between MIBC and NMIBC,further reflect the muscle invasiveness of bladder cancer.Therefore,the proposed radiomic strategy can be used to facilitate the preoperative prediction of muscular invasiveness in patients with bladder cancer.(3)This paper proposes multi-modal MRI based radiomic strategy for muscular invasiveness differentiation,and verify that features derived from the blend region of tumor & wall tissues have more potential to reflect the heterogeneous differences between MIBC and NMIBC than that from the convex tumor region.Results suggest that the feature selection strategy which included statistical analysis and RFE-SVM,and synthetic minority oversampling technique(SMOTE)for sample balance and augmentation,can further enhance the effect of invasiveness differentiation and stage prediction,and the radiomic features derived from multi-modal MR image and the blend region of tumor & wall tissues have more potential for cancer invasiveness differentiation.(4)This paper proposes PG-CMF BC segmentation strategy,realizes the accurate segmentation of BC.The results indicate that PG-CMF can be used for accurate and efficient segmentation of BC(5)This paper proposes the concept of DRI and its calculation algorithm to quantitatively evaluate the invasion degree of cancerous tissues into the bladder wall.Based on the DRI of BC,we further propose the quantitatively staging criteria for CADx of BC stage.