Application Of Magnetic Resonance Imaging-Based Machine-Learning In Precision Management Of Prostate Cancer

Objective: Prostate cancer(PCa)has become the most common urinary malignancy in men since 2008,according to China’s National Cancer Registry.Multiparametric magnetic resonance imaging(mp MRI)has become an important non-invasive technique for the detection of clinically significant prostate cancer(cs PCa).Should biopsies be delayed for patients with clinically suspected PCa but negative prostate mp MRI(PI-RADS≤2)? Risk stratification for patients with negative mp MRI can help reduce unnecessary systematic biopsies.Extended Pelvic node dissection(e PLND)is the most standard lymph node staging method.Not all patients with PCa have the same risk of lymph node invasion(LNI).To identify patients at low risk of LNI who can safely avoid unnecessary e PLND,several nomograms have been developed and validated to identify LNI.However,these nomograms are usually developed using clinical and biopsy parameters.Radiomic characteristics can provide important information about cancer characteristics,such as tumor shape,heterogeneity,genetic changes,and microenvironment,which may reflect tumor aggressiveness.The models based on traditional clinical factors and radiomic features may help predict biopsy results in mp MRI negative patients and preoperative prediction of LNI in PCa patients.The purpose of this study was to investigate the diagnostic value of radiomic features of mp MRI in predicting biopsy results in mp MRI negative patients and LNI in patients with PCa.Methods: 330 patients who underwent 3T mp MRI between January 2016 and December2018 and did not have PI-RADSV2 >3 lesions and underwent systematic biopsy within six months were enrolled in this study.Preselection of clinical features was performed using univariate analysis(p<0.05).Manually delineate the whole prostate gland ROI layer by layer to form VOI.Three hundred radiomic features were extracted from T2-weighted images and ADC images of the whole prostate gland and standardized by Z-Score.Sequential Floating Forward Selection algorithm(SFFS)is used for feature Selection.SVM classifier was used to model training to predict negative prostate biopsy results.The model performance was verified by keep-one cross-validation(LOOCV).The models were evaluated by AUC,sensitivity,specificity,and negative predictive value.Compare AUCs of different models by De Long test.In the study of predicting lymph node metastasis in patients with PCa,244 patients who underwent radical prostatectomy and extended pelvic lymph node dissection between January 2010 and December 2019 and who underwent mp MRI within six months before surgery were enrolled in this study.Patients were divided into training set(2010-2016)and test set(2017-2019)according to the date of MRI examination.In the training set,univariate logistic regression analysis was used to preselect clinical features.Manually delineate the ROI of prostate index lesions layer by layer to form VOI.220 radiomic features were extracted from T2-weighted images and ADC images of index lesions VOI and standardized by Z-Score.SFFS is used for feature selection.Model training using SVM classifier based on radiomics and clinical features to predict the outcome of pelvic lymph node metastasis.The model performance is verified in the validation group.All models were evaluated by AUC,sensitivity,specificity,and Negative predictive Value.Compare AUCs of different models by De Long test.Results: 1.Comparison of demographic data and clinical characteristics The median age of the 330 patients included was 63 years(IQR 58-67).Among 330 patients,306 had negative prostate systematic biopsy,and the NPV of mp MRI for cs PCa was 92.7%.Compared with biopsy-negative patients,patients with biopsy-proven cs PCa were older,had higher prostate specific antigen density(PASD),and smaller prostate volume(PV)(p<0.05).The median age of the 244 patients included was 63 years(IQR 59-67).17(10.6%)and 14(16.7%)patients with pelvic lymph node metastasis were included in the training and testing sets,respectively.Compared with patients without pelvic lymph node metastases,patients with pelvic lymph node metastases had higher PSA,PSAD,biopsy Gleason score and percentage of positive biopsy cores,tumor MR stage and more peripheral zone involvement(all p<0.05).2.Model building and evaluationIn the study predicting prostate biopsy result in mp MRI-negative patients,a total of 9features,including 6 radiomics and 3 clinical features,were included in the final SVM model.The combined machine learning model(AUC=0.798)had the highest AUC compared to models built using radiomics alone(AUC=0.679)or clinical features(AUC=0.705,p=0.011,0.006).The comprehensive machine learning model NPV of 98.3%is slightly higher than the NPV(93.9%)based on PSAD risk assessment.In the study of predicting pelvic lymph node metastasis in prostate cancer patients,among the radiomic features and pre-selected clinical features,the SFFS algorithm selected a total of 11 features into the SVM model,including 1 laboratory test feature,1 physical examination feature,3 MRI-observable features,3 biopsy-related features,and 3 radiomics features.In the test set,the combined machine learning model(AUC=0.915)had the highest AUC compared to models built using radiomics alone(AUC=0.6843)or clinical features(AUC=0.730,p=0.007,0.006).In the test set,the AUC of the integrative model established in this study reached 0.915(95% CI: 0.846-0.984),which was superior to existing clinical nomograms(AUC ranged from 0.698 to 0.724,p<0.05).Conclusion: Radiomics signatures can help predict biopsy results in patients with negative mp MRI and pelvic lymph node metastases in prostate cancer patients.The integrative models based on radiomics,and clinical features has the potential to predict biopsy results in patients with negative mp MRI and the risk of pelvic lymph node metastasis in prostate cancer patients,and guide the precise diagnosis and treatment of PCa,and safely avoid unnecessary prostate biopsy and pelvic lymph node dissection.Radiomics analysis has a promising application prospect in the precise diagnosis and treatment of prostate cancer.