Automated Contour Mapping With a Regional Deformable Model

Purpose

To develop a regional narrow-band algorithm to auto-propagate the contour surface of a region of interest (ROI) from one phase to other phases of four-dimensional computed tomography (4D-CT).

Methods and Materials

The ROI contours were manually delineated on a selected phase of 4D-CT. A narrow band encompassing the ROI boundary was created on the image and used as a compact representation of the ROI surface. A BSpline deformable registration was performed to map the band to other phases. A Mattes mutual information was used as the metric function, and the limited memory Broyden-Fletcher-Goldfarb-Shanno algorithm was used to optimize the function. After registration the deformation field was extracted and used to transform the manual contours to other phases. Bidirectional contour mapping was introduced to evaluate the proposed technique. The new algorithm was tested on synthetic images and applied to 4D-CT images of 4 thoracic patients and a head-and-neck Cone-beam CT case.

Results

Application of the algorithm to synthetic images and Cone-beam CT images indicates that an accuracy of 1.0 mm is achievable and that 4D-CT images show a spatial accuracy better than 1.5 mm for ROI mappings between adjacent phases, and 3 mm in opposite-phase mapping. Compared with whole image–based calculations, the computation was an order of magnitude more efficient, in addition to the much-reduced computer memory consumption.

Conclusions

A narrow-band model is an efficient way for contour mapping and should find widespread application in future 4D treatment planning.

Deformable model, Image registration, Contour mapping, IGRT