ConStruct is a software tool for image analysis in radiation therapy under development at UNC, expected to be released in spring 2008. It brings together research software for automatic and manual segmentation, and deformable and rigid registration, into a single usable package. The tool is designed with adaptive and image-guided radiation therapy in mind. For manual segmentation, it extends functionality found in in the treatment planning system PlanUNC. For automatic segmentation, the new tool permits two approaches. One approach relies on fitting a model called an m-rep to an image, and the other relies on deformable image registration.
The image analysis challenges of ART and IGRT stem from the large number of images involved. An essential step in radiation therapy is accurate three-dimensional segmentation of the target and organs at risk. Traditionally, this has been done by manual contouring on a slice-by-slice basis, but ART and IGRT lead to the acquisition of many treatment-time images per patient, which one may want to segment.
Also, image registration may be used to align the target with its location in the planning image, and deformable registration can be used to place dose distributions from different days in a common “tissue-based” frame of reference so that the cumulative delivered dose can be assessed.
|
|
|
| Main and 3D windows |
For manual segmentation, the tool contains a full suite of contour editing features. For instance, the user can draw contours, edit them by moving or cutting out points, copy and move them, and delete them. These facilities are based on the contouring component of the free research treatment planning system PlanUNC. ConStruct natively reads and writes PlanUNC contours and images, and also has software to read and write DICOM and DICOM-RT data.
The segmentation window is also used to compare multiple registered images, along with corresponding contours. This is important if consistency is desired between segmentations on different days. Two images can be viewed simultaneously via a blend slider, and a series of images can be shown in the form of a cine loop, along with the corresponding segmentations.
M-rep segmentation is a Bayesian approach in which a particular kind of generic 3D model of an organ is fit to an image using an optimization process. The optimization takes into account the plausibility of the shape the model is being asked to attain, and the degree to which the image intensities in the region around the model match those that would be expected near corresponding parts of the given organ.
The particular type of 3D model used is a medial representation, or m-rep. Discrete samples along a surface through the middle of the object are specified, and ``spokes'' point symmetrically outward from each sample point, the endpoints of the spokes marking the boundary.
M-rep model of a bladder, prostate, and rectum.
The deformable registration algorithm deforms the image so as to minimize the mean squared intensity difference between corresponding voxels, subject to a penalty for the degree to which the deformation differs from the identity. To allow large deformations, a time parameter is included, and the deformation penalty is based on the velocity of change, rather than the absolute magnitude. The algorithm progresses by calculating image forces based on the gradient of the moving image and on the intensity differences, and simulates the flow of a highly viscous compressible fluid, carrying the image into better alignment. The resulting deformations can be used to deform segmentations generated on one image to another, and they can also be applied to dose distributions calculated using other tools.
|
|
| Daily image with contours overlayed from planning image. | Daily image warped to fit planning image, with the same contours from planning image. |
The segmentation and registration tool provides a consistent user interface to apply and compare multiple segmentation methods to a single image, and across multiple images. It is being provided to the research community free of charge as a stand-alone tool intended to interact with standard DICOM RTP systems.
This poster presentation gives a brief overview of the project. It is a single PowerPoint slide.
These two papers discuss the deformable registration methodology used. There is also an introductory web page for this work.
The following paper discusses m-rep segmentation in the context of radiation therapy.
Research supported by NCI/NCRR R01 RR 01861501.