Automatic Bone and Marrow Extraction from Dual Energy CT through SVM Margin-Based Multi-Material Decomposition Model Selection

In this work, we present a fully-automatic approach for segmenting bone and marrow structures from dual energy CT (DECT) images. The images are represented using a multi-material decomposition model (MMD) computed from a triplet of physical materials at two different energy attenuation levels. We employ support vector machine learning to select the most relevant MMD model for the anatomical structure of interest so that highly accurate segmentation of the said structures can be achieved. We evaluated our approach for segmenting bone and marrow structures with varying amounts of metastatic bone disease on multiple longitudinal follow up patient scans. Our approach shows consistent and robust segmentation despite changes in bone density due to disease progression, high-density contrast material uptake in neighboring tissue, and significant metal artifacts.