Shape Modeling and Atlas-Based Segmentation for Identification of Lower Leg Tissues in pQCT

Abstract

In this work, we introduce an atlas-based segmentation method for lower leg tissues at 4%, 38%, and 66% tibial length. Our goal is to model the shape of the lower leg tissue types and to identify hard and soft tissues in an automated way. In our methodology, we implemented B-spline based free form deformation (FFD), and symmetric diffeomorphic demons (SDD) deformable models for nonlinear registration, and compared their performances for atlas-based segmentation accuracy on our pQCT data. Overall, we concluded that atlas-based segmentation is a promising technique, especially in the presence of noise and other types of image degradation. We also observed that the diffeomorphic demons algorithm may produce more accurate deformation fields than FFD. On the other hand, FFD produced smoother deformations than SDD. Quantitative analysis using the Dice similarity coefficient (DSC), showed that FFD was slightly better than SDD in identification of the trabecular bone tissue in 4% tibia. At 38% tibial length, SDD produced consistently higher DSC values than FFD, while at 66% tibia, FFD produced slightly higher segmentation accuracy.