Automatic Hippocampus Segmentation of Magnetic Resonance Imaging Images Using Multiple Atlases

摘要

The approach of hippocampus segmentation based on multiple atlases transfers segmentations from atlases to a new image by employing a non-rigid registration method. In this work, a multi-atlas-based automatic approach for segmenting the hippocampus in magnetic resonance imaging (MRI) is presented. In this approach, all the atlases are transformed to the query image firstly by applying Deformable Registration via Attribute Matching and Mutual-Saliency weighting (DRAMMS) to get deformation fields and warped atlas gray images. Then, by manipulating atlases and utilizing deformation fields, warped labels are achieved as a result of preliminary segmentation results. Finally, the hippocampus label images from all the atlases are further fused since each atlas contribute in the previous step. In considering the hippocampus, a relatively small area of the brain, certain strategies can be used to improve computational efficiency. The method of multi-atlas-based automatic segmenting is quantitatively validated by experiments on dataset from MICCAI 2012; experimental results suggest good segmentation accuracy and automation for clinical application.