A deformable-model approach to semi-automatic segmentation of CT images demonstrated by application to the spinal canal
- 22 January 2004
- journal article
- Published by Wiley in Medical Physics
- Vol. 31 (2), 251-263
- https://doi.org/10.1118/1.1634483
Abstract
Because of the importance of accurately defining the target in radiation treatment planning, we have developed a deformable-template algorithm for the semi-automatic delineation of normal tissue structures on computed tomography (CT) images. We illustrate the method by applying it to the spinal canal. Segmentation is performed in three steps: (a) partial delineation of the anatomic structure is obtained by wavelet-based edge detection; (b) a deformable-model template is fitted to the edge set by chamfer matching; and (c) the template is relaxed away from its original shape into its final position. Appropriately chosen ranges for the model parameters limit the deformations of the template, accounting for interpatient variability. Our approach differs from those used in other deformable models in that it does not inherently require the modeling of forces. Instead, the spinal canal was modeled using Fourier descriptors derived from four sets of manually drawn contours. Segmentation was carried out, without manual intervention, on five CT data sets and the algorithm's performance was judged subjectively by two radiation oncologists. Two assessments were considered: in the first, segmentation on a random selection of 100 axial CT images was compared with the corresponding contours drawn manually by one of six dosimetrists, also chosen randomly; in the second assessment, the segmentation of each image in the five evaluable CT sets (a total of 557 axial images) was rated as either successful, unsuccessful, or requiring further editing. Contours generated by the algorithm were more likely than manually drawn contours to be considered acceptable by the oncologists. The mean proportions of acceptable contours were 93% (automatic) and 69% (manual). Automatic delineation of the spinal canal was deemed to be successful on 91% of the images, unsuccessful on 2% of the images, and requiring further editing on 7% of the images. Our deformable template algorithm thus gives a robust segmentation of the spinal canal on CT images. The method can be extended to other structures, although it remains to be shown that chamfer matching is sufficiently robust for the delineation of soft-tissue structures surrounded by soft tissue.Keywords
This publication has 24 references indexed in Scilit:
- Assessment of consistency in contouring of normal‐tissue anatomic structuresJournal of Applied Clinical Medical Physics, 2003
- Rectum contouring variability in patients treated for prostate cancer: impact on rectum dose–volume histograms and normal tissue complication probabilityRadiotherapy and Oncology, 2002
- Loop-free snakes for highly irregular object shapesPattern Recognition Letters, 2002
- A review of deformable surfaces: topology, geometry and deformationImage and Vision Computing, 2001
- Intra- and inter-observer variability in contouring prostate and seminal vesicles: implications for conformal treatment planningRadiotherapy and Oncology, 1998
- Characterization of signals from multiscale edgesIEEE Transactions on Pattern Analysis and Machine Intelligence, 1992
- A theory for multiresolution signal decomposition: the wavelet representationIEEE Transactions on Pattern Analysis and Machine Intelligence, 1989
- Snakes: Active contour modelsInternational Journal of Computer Vision, 1988
- Hierarchical chamfer matching: a parametric edge matching algorithmIEEE Transactions on Pattern Analysis and Machine Intelligence, 1988
- Automatic Outlining of Regions on CT ScansJournal of Computer Assisted Tomography, 1981