CSAIL Event Calendar: Previous Series

Non Linear Registration Methods for Atlas-Based Segmentation of Anatomical Structures for Radiotherapy

Speaker: Olivier Commowick , INRIA/Children's Hospital
Date: April 27 2007
Time: 3:00PM to 4:00PM
Location: 32-D507
Host: Polina Golland, CSAIL

Radiotherapy planning requires the delineation of both the tumor and the organs at
risk. These segmentations are then used to optimize the doses of irradiation received on
each organ. This segmentation process is usually done manually and is therefore tedious
and not reproducible.

To solve those problems, we use an anatomical atlas, i.e. a representative anatomy
associated to a clinical image representing it. The registration of this atlas onto a
given patient allows then to segment automatically the patient structures. In this
approach, the registration method used to perform this step is crucial to get a good
segmentation result.

For our particular application, we indeed want the registration method to be as
independent as possible with respect to the setting of its parameters. This setting, done
by the clinician, indeed needs to be minimal (i.e. using one parameter set for all
patients) while guaranteeing a robust and smooth result on images coming from different
centers. To this end, three registration methods aiming at a better control the obtained
transformation will be presented:

- first, the introduction of deformability statistics (either scalar or tensor-based) in
the regularization of an existing registration algorithm will be presented, allowing to
increase the regularity of the obtained transformation,

- then, a new registration algorithm based on a Block-Matching technique will be
presented. A dense transformation is then interpolted between the pairings and thanks to
the use of an outlier rejection, this algorithm is very robust and produce smooth
deformations,

- finally, a locally affine registration method will be presented. This method is used to
register structures predefined on the atlas using local affine transformations. The
global transformation is interpolated at the end using the Log-Euclidean polyaffine
framework to ensure a smooth and invertible transformation. This method increases even
more the smoothness and the robustness of the registration as the number of degrees of
freedom of the transformation is reduced.

Segmentation results and the evaluation of those registration methods will be presented on
two applications:

- atlas-based segmentation of brain critical structures,
- construction of a head and neck atlas from a dataset of images.

See other events that are part of Biomedical Imaging and Analysis 2006/2007

See other events happening in April 2007