Automatic Segmentation and Discrimination of Connected Joint Bones from CT by Multi-atlas Registration

Abstract

Many applications require the automatic identification of bone structures in CT scans. The segmentation of the bone at the joints, however, is a difficult task to automate since the separation of the bones can be reduced by a degradation of the articular cartilage. In addition, the bone boundary can become very thin at certain locations due to osteoporosis, making it difficult to discriminate between the bone and neighbouring soft tissue. In this work, therefore, a probabilistic method is proposed to segment the bone structures by the registration of multiple atlases. Several atlas combination strategies are evaluated with respect to the segmentation and discrimination of the proximal femur and pelvic bone, and the L2 and L3 vertebrae, on datasets of 30 subjects using a leave-one-out approach. The mean overlap is computed and a false overlap measure is proposed to assess the correct discrimination of the bone structures. In addition, the mean average surface distances and Hausdorff distances are computed on the surface meshes extracted from the label maps. The results indicate that a generalized local-weighted voting approach is preferred, which results in a mean overlap $\ge 0.97$ for all bone structures, while being able to accurately discriminate between neighbouring bone structures.