3D Convolutional Neural Network for Segmentation of the Urethra in Volumetric Ultrasound of the Pelvic Floor

Pelvic organ prolapse (POP) decreases the quality of life for many women. To assess POP, the levator hiatus is segmented in a 2D plane of minimal hiatal dimensions, known as the C-plane. In order to automate plane detection, landmark information of key structures should be given to a plane detection algorithm. In this work, we present a fully automatic method to segment the urethra from a 3D transperineal ultrasound volume using a convolutional neural network (CNN). A dataset with 35 volumes from 20 patients during the Valsalva manoeuver (i.e. Valsalva, contraction and rest) labelled by an expert, was used for training and evaluation in a 5-fold cross-validation process. The 3D CNN model yielded an average robust Hausdorff distance of 4.68mm (95 percentile) which was comparable to intra-observer results.