Segmentation of the Biological Rat Spinal Cord Phantom from Diffusion Tensor Magnetic Resonance Image with a Nonparametric Statistical Front Evolution Method

The main problem of the most previous works on DTI segmentation is the parametric approach used for modeling tensor statistics in a region of interest which may not effectively model the tensor statistics in white-matter fiber bundles. In this work, we propose the Parzen density estimation with a Gaussian kernel to define region statistics in diffusion tensor images. This estimation is used to segment biological rat spinal cord phantom in a nonparametric statistical surface evolution algorithm framework. We demonstrate through numerical experiment that the statistical evolution surface with Euclidean distance derived from the non-parametric approach outperforms the evolution surface associated to the parametric case. In fact, both statistical modeling and metric are important to improve the quality of DTI segmentation results. Finally, it has shown that the most important element of kernel density estimation is the choice of the bandwidths. We conclude that if the parametric statistical modeling does not end up in a specific application with desired segmentation results, despite high computation cost of nonparametric method, nonparametric method is a reasonable choice.