A Primary Study Of Radiomic Multi-parametric Magnetic Resonance Imaging In The Diagnosis Of Prostate Cancer

Objective: This study was to investigate the value of four machine learning models based on DWI,T2 WI,DCE-early enhanced early and DCE-late enhanced advanced models in the diagnosis of prostate cancer.Materials and Methods: We retrospectively collected 127 cases of prostate cancer and 265 cases of benign prostatic hyperplasia who had MRI examination and confirmed by pathology in our hospital from January 1,2015 to June 30,2017.All patients included complete axial T2 WI,DWI and LAVA Sequence DCE Dynamic Enhancement Sequence.Patient data was exported from the PACS system(Picture Archiving and Communication Systems)using the AW4.6 workstation,including full T2 WI horizontal axis,DWI axis,and DCE axis images.Understanding the histopathological findings,reviewed each dataset and two radiologists with more than five years of work experience identified the most suspicious locations for prostate cancer based on histopathology,selecting the area with the largest lesion size and determining the extent of the lesion boundary,If multiple lesions,take the largest lesions.If there is any disagreement,it is necessary to reach a consensus through consensus.In the AK software,we selected the level of the largest cross-section of prostate cancer determined by histological-MRI matching and manually outlined ROI along the boundary of the lesion.Prostate hyperplasia cases select the largest cross-sectional area of the prostate,manually outline the region of interest along the border of the prostate(ROI).At the same time try to keep the drawing line at a distance of about 1-2 mm from the edge of the lesion,so as to minimize the partial volume effect of the edge.396 texture features were calculated.Run the Correlation Matrix through the 'corrplot' package to remove redundant features.Through the 'Boruta' feature extraction method,32 features were extracted by DWI,24 features were extracted by T2 WI,26 features were extracted early by DCE,and 17 features were extracted by late DCE.Using the caret package in the R software(version 3.4.2),By extracting features we set up four machine learning models of liner discriminant analysis(LDA),Random Forest(RF),K-nearest neighbor(K-NN)and Naive Bayesian(NB),based on DWI,T2 WI,DCE enhanced early,DCE enhanced late,respectively.The 127 cases of prostate cancer and 265 cases of benign prostatic hyperplasia were randomly divided into training set and test set according to the ratio of 70% to 30%.The training set consists of 89 cases of prostate cancer and 186 cases of benign prostatic hyperplasia.The validation set consists of 38 cases of prostate cancer and 79 cases of benign prostatic hyperplasia.The training set is divided into ten groups at random,and the ten-fold cross-validation method is used to train the classifier to obtain the machine learning model.The test set was used to verify the accuracy,sensitivity,specificity,ROC curve,area under the ROC curve(AUC)and 95% confidence interval.Results: The four machine learning AUC values show the higher diagnostic performance from the lowest 0.747 to the highest 0.9327.In the same machine learning model based on texture parameters of different series,the AUC values of machine learning model based on DWI are higher than others,Especially the AUC value of the RF machine learning model based on DWI is up to 0.9327,achieving excellent classification performance.The AUC values of the RF models in the four machine learning models are higher than those of the LDA,K-NN and NB models.Conclusion: The machine learning model was established based on the radiomic characteristics of multi-parameter magnetic resonance imaging(mp MRI),which has good diagnostic performance for differentiating prostate cancer and hyperplasia of prostate.It has great potential and clinical significance to differentiate prostate hyperplasia and prostate cancer,and it may become an auxiliary diagnostic tool for radiologists in the future.