MR prior based automatic segmentation of the prostate in TRUS images for MR/TRUS data fusion

The poor signal-to-noise ratio in transrectal ultrasound (TRUS) images makes the fully automatic segmentation of the prostate challenging and most approaches proposed in the literature still lack robustness and accuracy. However, it is relatively straightforward to obtain high quality segmentations in magnetic resonance (MR) images. In the context of MR to TRUS data fusion the information gathered in the MR images can hence provide a strong prior for US segmentation. In this paper, we describe a method to non-linearly register a patient specific mesh of the prostate build from MR images to TRUS volume. The MR prior provides shape and volume constraints that are used to guide the MR-to-TRUS surface deformation, in collaboration with a US image contour appearance model. The anatomical point correspondences between the MR and TRUS surfaces are obtained implicitly. The method was validated on 30 pairs of MR/TRUS patient exams and achieves a mean Dice value 0.85 and a mean surface error of 2.0 mm.