Meeting Banner
Abstract #1509

Machine Learning for Target Selection in MR-Guided Prostate Biopsy: A Preliminary Study

Mehdi Moradi1, Andriy Fedorov1, William M. Wells1, Kemal Tuncali1, Sandeep N. Gupta2, Fiona M. Fennessy1, Clare M. Tempany1

1Radiology, Brigham and Women's Hospital - A Teaching Affiliate of Harvard Medical School, Boston, MA, United States; 2GE Global Research Center, Niskayuna, NY

We propose to use machine learning to enhance the process of target selection for 3T MR-guided transperineal prostate biopsy. Support vector machine and Gaussian classifiers with different combinations of diffusion and DCE MRI are examined. Training is performed on data from 13 prostatectomy cases with histologically confirmed cancer in the peripheral zone. The trained classifier was used to determine the outcome of in ten PZ biopsy samples from five patients. The Bayesian classifier with ADC as the only feature resulted in the ROC area of 0.964 in leave-one-patient-out cross-validation on the training dataset. The outcomes of eight of the ten biopsies, including all three cancer samples, were correctly determined.

Abstract PDF Poster PDF

Keywords

accuracy accurate achieved acknowledgments adding agree aided among annual approaches approximate assess assist assumed available bell benign best beyond biopsies biopsy cancer cancerous characteristics characterization class classification classified classifier close coil collected combinations common comprised compute computed confirm confirmed considered contoured correctly count cross curve dataset death detection diffusion distinctly distributions dynamic employed enable enrolled evaluation exams explains extend feasibility feature features finding findings five gadolinium global guided highly idea identified improve included larger learning least leave likelihood likelihoods limitations localization machine majority malignancies maps materials measure medical methodology move operative outcome outcomes patient patients peripheral pixel posterior predict preliminary priors probability process proposed prostate protocol radical radiologists remains removing report reported retrospectively review routine rule sample samples scanner school selected selection sensitivity separation seven shaped solution stage stages strong supervised suspicious table target targeting targets teaching towards train training tumors undergone underwent validated validation vector wash women yield yielded zone