The Preliminary Research Of MRI Radiomics In Diagnosis Of Prostate Cancer And The Prediction Of TMPRSS2-ETS Fusion Gene In Prostate Cancer

Part 1 Value of bp-MRI radiomics and ADC value in the diagnosis of prostate transition zone and peripheral zone cancer 0.927(readerl),0.807,0.857,0.879(reader2)and 0.864,0.890,0.928(reader2).By comparing the diagnostic efficacy of different b values in the diagnosis of transitional zone carcinoma of prostate,the diagnostic efficiency of b=3000 s/mm2 was the maximum.And the diagnostic efficiency of b=3000 s/mm2 was also the maximum diagnosing peripheral prostate cancer.After five-fold cross-validation,the average test data showed that the AUC value of ultra-high b value(b=2000 and 3000 s/mm2)was the maximum.And the AUC values for the diagnosis of prostate transitional zone carcinoma were 0.81 56 and 0.8154,respectively.The AUC values for the diagnosis of prostate peripheral zone cancer were 0.8316 and 0.8427,respectively.Purpose:To investigate the value of ADC value of DWI in the diagnosis of prostate transitional zone and peripheral region cancer and to study the diagnostic value of bp-MRI in prostate transitional zone and peripheral region cancer.Materiala and Methods:Between October 2015 and December 2018,633 patients with prostate diseases who met the selection criteria participated in the study.The images of all patients were scanned on GE 3.0T MRI.The image sequences included transverse high resolution T2WI and DWI(b=0,1000,2000,3000 s/mm2).Pathological results were obtained in all cases.The ADC values of prostate cancer lesions and non-cancer lesions were measured by two experienced abdominal MRI diagnostic physicians and consistency and repeatability analysis were done.And the ADC values between prostate cancer lesions and non-cancer lesions were compared.The value of ADC value in the diagnosis of prostate transitional zone and peripheral area cancer was evaluated by ROC curve.After discussion by two experienced abdominal MRI diagnostic physicians,prostate cancer lesions and non-cancer lesions were manually labeled,and then 106 dimensional features were extracted using Radiomics.XGboost version 0.81 and Python version 3.6 were used as development tools to build the model.The dimensional features were reduced by PCA as input and the test set was input to obtain the prediction ability of the model.Finally,5-fold cross-validation was used to improve the accuracy of prediction and the generalization ability of the model.Results:The consistency and repeatability of ADC values measured by two radiologists between prostate cancer lesions and non-cancer lesions were excellent.There were significant differences in ADC values between transitional zone cancer and non-cancer lesions and between peripheral prostate cancer and non-cancer lesions(P<0.05).The area under the ROC curve of different b values for diagnotic prostate transitional zone cancer and peripheral cancer were respectively 0.798,0.857,0.884(reader 1),0.869,0.879,Conclusions:Ultra-high b-value DWI was of great value in the diagnosis of prostate cancer,and DWI of b=3000s/mm2 is the most valuable.Machine learning of bp-MRI radiomics can accurately and automatically distinguish prostate cancer from non-cancer lesion in different regions.Part 2 Diagnostic value of bp-MRI radiomics for Gleason classification of prostate cancerPurpose:To study high and low risk groups of Gleason score of prostate cancer using the ADC value and to assess the performance of neural network classification to stratify the Gleason Score of prostate cancer based on vast image features across bp-MRI.Materiala and Methods:316 cases of prostate cancer who met the selection criteria between October 2015 and December 2018 were collected in the study.MRI sequences included high-resolution transverse T2WI and DWI(b=0,1000,2000 and 3000 s/mm2).Surgical pathological results were obtained in all cases.The ADC values of prostate cancer lesions were measured by two experienced abdominal MRI diagnostic physicians and consistency and repeatability analysis were done.The ADC values of prostate cancer were compared between high-risk group and low-risk group and ROC curve was used to evaluate the value of ADC value of different b values in distinguishing high risk group from low risk group of Gleason of prostate cancer.Prostate cancer lesions were manually labeled after discussion by two experienced prostate diagnostic physicians.And then 106 dimensional features were extracted using Radiomics.Spearman nonparametric correlation test was carried out,and then dimensionality reduction was carried out.The constructed neural network was trained by prostate cancer as input,and the test set was input to obtain the prediction ability of the model.Finally,10-fold cross-validation and 100 times shuffle were used to improve the accuracy of prediction and the generalization ability of the model.Results:The consistency and repeatability of ADC values measured by two physicians in high and low risk groups of Gleason of prostate cancer were excellent.There was significant difference in ADC values of different b values between high risk group and low risk group of Gleason of prostate cancer(P<0.05).The area under the ROC curve of different b values distinguishing high risk group from low risk group of Gleason of prostate cancer were 0.658,0.662,0.683(readerl)and 0.671,0.660,0.686(reader2),respectively.By comparison,there was no significant difference in diagnostic efficacy between different b values.77-dimensional features had significant correlation at the p=0.05(double tail)level.After dimensionality reduction 21-dimensional new features were obtained which preserves 99%of the original feature information.The results of 10-fold cross-validation and shuffling validation data:T2WI sequence ACC=0.678,AUC=0.712,DWI(b=1000 s/mm2)ACC=0.662,AUC=0.689,DWI(b=2000 s/mm2)ACC=0.694,AUC=0.689 and DWI(b=3000 s/mm2)ACC=0.674,AUC=0.691.The diagnostic performance of T2WI was slightly higher than that of DWI.The comparison of AUC values of DWI sequences with different b values showed no significant difference.Conclusions:ADC values with different b value had certain value in differentiating high risk group from low risk group of Gleason of prostate cancer,but the diagnostic efficiency was lower than that of radiomics.Features extracting from bp-MRI and using neural network can accurately and automatically distinguish high-risk group from low-risk group gleason of prostate cancer.This model has strong ability of generalization and was not easy to over-fit.Part 3 Predictive value of bp-MRI imaging radiomics for T staging of prostate cancerPurpose:To investigate the relationship between ADC value and high risk group and low risk group of T staging of prostate cancer and to evaluate high riskgroup and low risk group of T stage of prostate cancer using bp-MRI radiomics.Materiala and Methods:217 eligible prostate cancer patients from January 2016 to April 2018 were enrolled in the study.The images of all patients were obtained on GE 3.0T MRI.The image sequences included axial high resolution T2WI and DWI(b=0,1000,2000 and 3000 s/mm2).The postoperative pathological results were obtained in all cases.The ADC values of prostate cancer lesions were measured by two prostate diagnostic physicians with more than ten years experience and consistency and repeatability analysis were done.The ADC values between high risk group and low risk group of T stage of prostate cancer were compared.The correlation between ADC values with different b values and high and low risk groups of T stage of prostate cancer was evaluated by logistic regression analysis.Prostate cancer lesions were manually labeled after discussion by two experienced prostate diagnostic physicians.And then 106 dimensional features were extracted using Radiomics.XGboost 0.81 version and Python3.6 version were used as development tools to build the model.The dimensional features were reduced by PC A as input and the test set was input to obtain the prediction ability of the model.Finally,5-fold cross-validation was used to improve the accuracy of prediction and the generalization ability of the model.Results:The consistency and repeatability of the ADC values of prostate cancer lesions measured by two physicians were excellent.And there was significant difference between high risk group and low risk group of T stage of prostate cancer(P<0.05).There was no significant correlation between ADC values with different b values and high and low risk groups of T stage of prostate cancer(P>0.05).The average test data after 5-fold cross-validation showed that the diagnostic efficiency of different sequence pairwise combinations was higher than that of each sequence.The AUC value of DWI(b=1000 s/mm2)+ DWI(b=2000 s/mm2)was the maximum and was 0.6862.Conclusions:There was no significant correlation between ADC value and high risk group and low risk group of T stage of prostate cancer.Machine learning of bp-MRI radiomics could distinguish high risk group from low risk group of T stage of prostate cancer.But the ability to automatically distinguish high risk group from low risk group of T stage of cancer need to be further improvedPart 4 The Predictive Value of biparametric MRI Imaging Radiomics in TMPRSS2-ETS Fusion Gene of Prostate CancerPurpose:To predict and analyze TMPRSS2-ETS fusion gene in prostate cancer by bp-MRI Radiomics.Materials and Methods:This study was a prospective study.Prostate cancer patients who met the requirements of this study were collected from November 2017 to February 2019.MRI-ultrasound fusion targeted biopsy was performed in all patients.After discussion by two experienced abdominal MRI radiologist,prostate cancer lesions were manually labeled.Then 354 dimensional features were extracted by Radiomics,and dimensionality reduction was carried out by variance threshold and SelectKBest.Then Spearman correlation analysis was used to analyze the correlation between the screened characteristics and fusion genes.Then the characteristics between fusion gene group and non-fusion gene group were compared,and ROC curve was used to analyze the diagnostic efficiency of different sequences for gene fusion group.Results:In this study 44 patients were collected,including 9 cases of fusion gene and 35 cases of non-fusion gene.After dimensionality reduction and correlation analysis,T2WI had 4-dimensional characteristics closely related with fusion gene.DWI(b=1000s/mm2,2000s/mm2 and 3000s/mm2)had 7-dimensional,6-dimensional and 7-dimensional characteristics closely related with fusion gene,respectively.23 of the 24-dimensional features related with fusion gene were significantly different between fusion gene group and non-fusion gene group(P<0.05).The AUC value of GrayLevelNonUniformity/glszm feature of T2WI is the maximum(0.813).The AUC value of GrayLevelNonUniformity/glszm feature of DWI(b=1000 s/mm2)is the maximum(0.844).The AUC value of RunLengthNonUniformity/glrlm feature of DWI(b=2000s/mm2)is the maximum(0.819)and the AUC value of SizeZoneNonUniformity/glszm feature of DWI(b=3000s/mm2)is the maximum(0.797).Conclusion:The feature values of bp-MRI imaging radiomics were different between TMPRSS2-ETS fusion gene and non-fusion gene group of prostate cancer.It was helpful to predict fusion gene.Radiomics can provide more information than traditional imaging.