LAUSR

available at http://pubmed.ncbi.nlm.nih.gov/31592731 Editorial Comment: In this retrospective study 312 men were evaluated with multiparametric magnetic resonance imaging and targeted transrectal fusion biopsy. T2-weighted and diffusion imaging sequences were used for PI-RADS (Prostate Imaging Reporting and Data System) clinical assessment by radiologists and for performance of U-Net, a neural network optimized for voxel based image segmentation. Implementing this U-Net system required extensive training, validation and testing. The study consisted of 250 cases for training and validation, and a test set of 62 cases. The purpose of this study was to compare U-Net performance to radiologist assessment with PI-RADS in patients with clinically significant prostate cancer (PI-RADS 4 or greater) as well as accuracy with segmentation of lesions and prostate boundaries. The authors found good agreement between computer generated and radiologist generated segmentation for the prostate boundary and lesion segmentation. They found the U-Net had similar performance to experienced prostate imagers for clinically significant prostate cancer. Overall, this study suggested that deep learning algorithms have the potential to aid the radiologist in identifying significant prostate cancer. The study had several limitations, including sample size and size of clinically significant prostate cancer lesions studied. How accurate are targeted biopsies in small lesions?