TU-A-9A-01: A Precise Deformable Image Registration System Using Feature-Based Irregular Meshes

Authors


Abstract

Purpose:

To develop a deformable image registration system using irregular meshes generated based on image features. Finer meshes are generated at organ or tissue boundaries so that higher deformation resolution is achieved with limited number of control points.

Methods:

This deformable image registration system consists of two parts: a finite element modeling system and a GPU-based registration system. The finite element modeling system is used to generate irregular meshes for the entire body or region of interest without segmentation. This system uses a Laplacian of Gaussion operator to extract features such as boundaries between organs and tissues. The placement of control points (vertices) of meshes is optimized based on feature intensities. The registration framework uses elastic energy as regularity and guarantees the final deformation vector field to be diffeomorphic. This system has been compared with registration algorithms based on regular meshes or voxels (Demons). The comparison is performed on XCAT digital phantom data and patient data.

Results:

Both XCAT male phantom and female phantom were used. Two respiratory phases were treated as floating and fix images for every phantom. Similarity measures were performed for image intensity and displacement vector field (DVF) for four sets of testing data with three types of deformation algorithms (irregular mesh, regular grid, and voxel). It clearly shows that the voxel-based Demons algorithm leads to slightly better images. However, the DVF results show that mesh-based algorithms behave much better. Compared with regular meshes, the irregular mesh we proposed leads to much faster convergence and converge to better results. The calculation is accelerated by GPU cards and a typical registration of patient data could be completed in about 1 minute.

Conclusion:

The feature-based irregular meshing method provides a fast and accurate deformable image registration.

This research is supported by CPRIT individual investigator award RP110329 and Varian research grant.

Ancillary