MO-C-17A-09: Deformation Image Registration Using a Spatial-Context Regularization Filter

Authors


Abstract

Purpose:

Intensity-based deformable registration with spatial-invariant regularization will generally fail when distinct motion exists across the interface of different type of tissues. The purpose of this work was to develop a new regularization scheme for deformable registration that is adaptive to motion discontinuities along different types of tissue.

Methods:

Our approach was built upon a multi-resolution Demons-type framework, and used the prior knowledge defined in the image context as an additional constraint to regularize deformation vector fields, which leads to a spatialcontext vector smoothing filter. This filter is spatial-variant and anisotropic, and can be reduced to the product of a Gaussian function of spatial locations and an additional Gaussian function of image intensity difference. This additional regularization favors the motion vectors within the same tissue type, but penalized the motions vectors from different tissues. Five lung patient cases from a benchmark data set, each with 300 landmark pairs delineated by a physician, were adopted for the validation of our algorithm. Our approach was also compared against a state-of-the-art implementation of dual-force Demons algorithm. We also included one difficult lung case with large sliding motion and one head-neck cases with large motion in air cavity into our validatio

Results:

For five lung cases, the mean and standard deviation of the landmark displacements were 1.3±0.8(mm) using our approach, but this number went up to 2.3±2.9(mm) using Demons algorithm. Particularly for a case with largest initial displacement of 15± 9(mm), our approach reduced the mean displacement to 1.3±1.1(mm), but Demons can only achieve 3.6±5.9(mm). We also found that, for the two difficult cases in which the original intensity-based registration failed the registration accuracy was drastically improved using the proposed algorithm.

Conclusion:

We have demonstrated the effectiveness of an algorithm using adaptive regularization in addressing large distinct motion across different types of tissue.

This work is partially supported by the CPRIT (Cancer Prevention Research Institute of Texas) grant RP110732

Ancillary