Multi-atlas and Gaussian Mixture Modeling Based Perirectal Fat Segmentation from CT Images

Abstract

Accurate perirectal fat segmentation in CT images aids in estimating radiation dose delivered to the region of fat around the rectum during radiation therapy treatment of prostate cancer. Such a process is important in determining the resulting toxicity of the neighboring tissues. However automatic or semi-automatic segmentation of the perirectal fat in CT images is a challenging task due to inter patient anatomical variability, contrast variability and imaging artifacts. We propose a combined schema of multi-atlas and multi parametric Gaussian mixture modeling for perirectal fat segmentation in CT images. Multi-atlas based soft segmentation and multi parametric Gaussian mixture modeling aids in identifying the volume of interest (VOI). Thereafter expectation maximization (EM) based soft clustering of the intensities of the VOI refined with positional probabilities of the perirectal fat provides the segmentation of the perirectal fat. The proposed method achieves a mean sensitivity value of 0.88±0.07 and a mean specificity value of 0.998±0.001 with 5 patient datasets in a leave-one-patient-out validation framework. Qualitative results show a good approximation of the perirectal fat volume compared to the ground truth.