PROST-Net: A Deep Learning Approach to Support Real-Time Fusion in Prostate Biopsy

Abstract

Prostate biopsy fusion systems employ manual segmentation of the prostate before the procedure, therefore the image registration is static. To pave the way for dynamic fusion, we introduce PROST-Net, a deep learning (DL) based method to segment the prostate in real-time. The algorithm works in three steps: firstly, it detects the presence of the prostate, secondly defines a region of interest around it, discharging other pixels of the image before the last step which is the segmentation. This approach reduces the amount of data to be processed during segmentation and allows to contour the prostate regardless of the image modality (e.g., magnetic resonance (MRI) or ultrasound (US)) and, in the case of US, regardless of the geometric disposition of the sensor array (e.g., linear or convex). PROST-Net produced a mean Dice similarity coefficient of 86% in US images and 77% in MRI images and outperformed other CNN-based techniques. PROST-Net is integrated in a robotic system–PROST– for trans-perineal fusion biopsy. The robot with PROST-Net gives the potential to track the prostate in real-time, thus reducing human errors during the biopsy procedure.