A fast model-based prostate boundary segmentation using normalized cross-correlation and representative patterns in ultrasound images

Abstract

Segmentation of prostate boundaries in transrectal ultrasound (TRUS) images plays a great role in early detection of prostate cancer. Due to the low signal to noise ratio and existence of the speckle noise in TRUS images, prostate image segmentation has proven to be an extremely difficult task. This paper introduces a new fully automatic model-based prostate boundary segmentation method based on normalized cross-correlation (NCC). Using lower and upper boundary representative patterns, a strip rotates around the center of the probe and emphasizes the prostate boundaries. Representative patterns are constructed from a dictionary learning method, referred to as iterative least squares dictionary learning algorithm (ILS-DLA). Affine transformation parameters transform the prostate model to a position that best fit on the emphasized boundaries. Dice similarity coefficient (DSC) is adopted to evaluate the accuracy of the automatic segmentation procedure. Successful experimental results and the average DSC value of 90.6% and computational time of 3.08 seconds validate the proposed method.