3D Multi-scale level set segmentation of vertebrae

Abstract

We present a 3D multi-scale segmentation algorithm for vertebrae based on a cascade of level sets. We validated the performance of the algorithm in terms of accuracy, robustness and speed and compared it with its original single-scale version using a synthetic vertebra and a semi-synthetic one containing natural structured noise. The multi-scale algorithm was found to be 4 to 5 times faster and could be half a voxel more accurate than the single-scale algorithm. Synthetic vertebrae also allowed us to try a large number of level set parameters. Analysis of the results revealed trends about parameter combinations leading to leakage or under segmentation. After selecting a good set of parameters we tested the algorithm on 50 real vertebrae. The success rates were 90% and 80% respectively for the multi-scale and single-scale algorithms.