This study aimed to investigate the performance of^(18)F-DCFPyL positron emission tomography/computerized tomography(PET/CT)models for predicting benign-vs-malignancy,high pathological grade(Gleason score>7),and cl...This study aimed to investigate the performance of^(18)F-DCFPyL positron emission tomography/computerized tomography(PET/CT)models for predicting benign-vs-malignancy,high pathological grade(Gleason score>7),and clinical D'Amico classifcation with machine learning.The study included 138 patients with treatment-naïve prostate cancer presenting positive^(18)F-DCFPyL scans.The primary lesions were delineated on PET images,followed by the extraction of tumor-to-backgroundbased general and higher-order textural features by applying fve diferent binning approaches.Three layer-machine learning approaches were used to identify relevant in vivo features and patient characteristics and their relative weights for predicting high-risk malignant disease.The weighted features were integrated and implemented to establish individual predictive models for malignancy(Mm),high path-risk lesions(by Gleason score)(Mgs),and high clinical risk disease(by amico)(Mamico).The established models were validated in a Monte Carlo cross-validation scheme.In patients with all primary prostate cancer,the highest areas under the curve for our models were calculated.The performance of established models as revealed by the Monte Carlo cross-validation presenting as the area under the receiver operator characteristic curve(AUC):0.97 for Mm,AUC:0.73 for Mgs,AUC:0.82 for Mamico.Our study demonstrated the clinical potential of^(18)F-DCFPyL PET/CT radiomics in distinguishing malignant from benign prostate tumors,and high-risk tumors,without biopsy sampling.And in vivo^(18)F-DCFPyL PET/CT can be considered a noninvasive tool for virtual biopsy for personalized treatment management.展开更多
文摘This study aimed to investigate the performance of^(18)F-DCFPyL positron emission tomography/computerized tomography(PET/CT)models for predicting benign-vs-malignancy,high pathological grade(Gleason score>7),and clinical D'Amico classifcation with machine learning.The study included 138 patients with treatment-naïve prostate cancer presenting positive^(18)F-DCFPyL scans.The primary lesions were delineated on PET images,followed by the extraction of tumor-to-backgroundbased general and higher-order textural features by applying fve diferent binning approaches.Three layer-machine learning approaches were used to identify relevant in vivo features and patient characteristics and their relative weights for predicting high-risk malignant disease.The weighted features were integrated and implemented to establish individual predictive models for malignancy(Mm),high path-risk lesions(by Gleason score)(Mgs),and high clinical risk disease(by amico)(Mamico).The established models were validated in a Monte Carlo cross-validation scheme.In patients with all primary prostate cancer,the highest areas under the curve for our models were calculated.The performance of established models as revealed by the Monte Carlo cross-validation presenting as the area under the receiver operator characteristic curve(AUC):0.97 for Mm,AUC:0.73 for Mgs,AUC:0.82 for Mamico.Our study demonstrated the clinical potential of^(18)F-DCFPyL PET/CT radiomics in distinguishing malignant from benign prostate tumors,and high-risk tumors,without biopsy sampling.And in vivo^(18)F-DCFPyL PET/CT can be considered a noninvasive tool for virtual biopsy for personalized treatment management.