Share Email Print

Proceedings Paper

Semi-synthetic digital phantoms incorporating natural structured noise and boundary inhomogeneities
Author(s): Sovira Tan; Michael M. Ward
Format Member Price Non-Member Price
PDF $17.00 $21.00

Paper Abstract

Validating segmentation algorithms remains a difficult problem. Manual segmentation taken as gold standard is timeconsuming and can still be contentious especially in the case of complex 3D objects and in the presence of important partial volume effect (PVE). In contrast digital phantoms have well-defined built-in boundaries even when PVE is simulated. However their degree of realism is questionable. In particular the rich natural structures inside an object that constitute one of the most difficult obstacles to segmentation are to this day too complex to model. A new method for constructing semi-synthetic digital phantoms was recently proposed that incorporates natural structured noise and boundary inhomogeneities. However only one phantom was presented and validation was lacking. In the present work we constructed 5 phantoms of vertebral bodies. Validation of phantoms should test their ability to predict how an algorithm will perform when confronted to real data. Our phantoms were used to compare the performance of two level set based segmentation algorithms and find the parameters that optimize their performances. We validated the phantoms by correlating the results obtained on them with those obtained on 50 real vertebrae. We show that: 1) the phantoms accurately predict which segmentation algorithm will perform better with real clinical data. 2) by combining the results obtained by the 5 different phantoms we can extract useful predictions about the performance of different sets of parameters on real data. Because the phantoms possess the high variability of real data predictions based on only one phantom would fail.

Paper Details

Date Published: 11 March 2008
PDF: 8 pages
Proc. SPIE 6914, Medical Imaging 2008: Image Processing, 69144W (11 March 2008); doi: 10.1117/12.770465
Show Author Affiliations
Sovira Tan, National Institutes of Health (United States)
Michael M. Ward, National Institutes of Health (United States)

Published in SPIE Proceedings Vol. 6914:
Medical Imaging 2008: Image Processing
Joseph M. Reinhardt; Josien P. W. Pluim, Editor(s)

© SPIE. Terms of Use
Back to Top
Sign in to read the full article
Create a free SPIE account to get access to
premium articles and original research
Forgot your username?