2014 Summer Project Week:Image Registration with Sliding Motion Constraints

From NAMIC Wiki
Revision as of 14:10, 27 June 2014 by Derksen (talk | contribs)
Jump to: navigation, search
Home < 2014 Summer Project Week:Image Registration with Sliding Motion Constraints

Key Investigators

Fraunhofer MEVIS: Alexander Derksen, Kanglin Chen

MGH: Gregory Sharp

MIT: Danielle Pace?

Project Description

Objective

Common deformable image registration approaches try to estimate a continuous and smooth deformation field. However in certain cases, e.g. registration of respiratory lung CT images, a globally continuous estimation of the deformation is not plausible, i.e. sliding motion of organs can not be described in a continuous way. Due to respiratory motion induced sliding of the lung along the neighboring tissue, an estimation that is discontinuous along the lung surface and continuous elsewhere would be expected and more plausible. In this project we work on a deformable image registration approach that takes sliding motion of organs into account, by explicitly modeling discontinuities in the deformation field.

Approach, Plan

  • The general registration approach is formulated as optimization problem.
  • Sliding motion of organs is explicitly modeled as constrained to the registration problem.
  • The problem is discretized with a FEM approach. A numerical solution is computed with a Gauß-Newton scheme.
  • A 2D implementation in Matlab exists.
  • TODO:
    • Extend code to work on 3D data
    • Rewrite (core) methods as mex files...
    • ...do it in a highly parallel form (OpenMP, CUDA, OpenCL,...?)
    • Test algorithm on lung CT scans

Progress

  • A 3D implementation with simplified Sliding Constraint was implemented
  • An inhale-exhale lung CT scan (DIR Lab) was registered and visually evaluated
  • Result with simplified constraint: Sliding behavior does occur at lung boundary. However significantly better results are expected from non-simplified implementation.
  • Good progress on non-simplified Sliding Constraint (Gradient and approximated Hessian for optimization implemented)