Automatic segmentation of prostate in MR images using deep learning and multi-atlas techniques

Abstract

Precise segmentation of prostate in magnetic resonance images is an essential step in treatment planning and a challenging task due to high variability in shape and size of the tissue. In this paper, we propose an automatic algorithm for accurate and robust segmentation of prostate in MR images. First, we employ a deep neural network to locate the prostate region of interest which removes background pixels and reduces the size of the image. Then, we obtain an initial segmentation of the tissue using a probabilistic atlas. Finally, we utilize statistical shape models to restrict the final contour inside the allowable shape domain. We performed a quantitative evaluation on 30 MR images and obtained a mean Dice similarity coefficient of 0.85±0.06. Compared to recent researches, our method is both robust and accurate.