RSNA 2003 Scientific Papers > Evaluation of the Performance of a Bone Segmentation ...
  Scientific Papers
  SESSION: Neuroradiology/Head and Neck (Cerebral Blood Flow I)

Evaluation of the Performance of a Bone Segmentation Algorithm for Head/Neck CTA

  DATE: Wednesday, December 03 2003
  START TIME: 11:10 AM
  END TIME: 11:17 AM
  LOCATION: Room N226
  CODE: K12-964

Srikanth Suryanarayananan PhD
Rakesh Mullick PhD
Vidya Kamath MD
Arjun Kalyanpur MD
Uday Patil MBBS, MD
Danielle Drummond

Computed tomography (CT), angiography
Head, CT
Images, processing


Purpose: Our goal is to evaluate the performance of a segmentation algorithm to remove bone from head/neck CT Angiography by 2 radiologists and 2 evaluation methods.

Methods and Materials: CT Angiography data sets of the head/neck region were acquired using different configurations of CT scanners (GE Medical Systems, Milwaukee, USA). The algorithm loads each case as a series of axial slices for processing. An automatic partition algorithm is applied to identify the skull base region and create separate sub-volumes. The sub-volumes are processed separately and merged together using a constrained region growing. The segmented models are reviewed by two different radiologists. The first evaluation is performed using static panels of the 3-D models displayed using volume rendering and MIP modes. A second evaluation is performed by the use of a VTK based interactive evaluation tool. The tool allows the user to scroll through the slices and select a suitable window/level setting. The original and the segmented data both be examined. The 3-D vessel model can be rotated and zoom in/out to allow the user a 360 degree viewing capability. All the radiologists use a set of questions as a guide to evaluate the algorithm performance and determine a rating. The rating is on a 1-5 scale, going from worst to best. While using the VTK evaluation tool, the questions are built into the GUI where the radiologist can click on the rating number. The mean rating is compared across the radiologists using the static panels and the 3-D tool.

Results: The overall mean rating (1: worst, 5: best) of this bone segmentation algorithm by radiologist A was 3.52 (Std. Dev. 0.68) using panels of 2-D VR and MIP images and rated 3.13 (Std. Dev. 0.42) using the interactive 3-D evaluation tool. The ratings from radiologist B using static 2-D panels was 3.56 (Std. Dev. 1.27) and the ratings from using the 3-D evaluation tool was 3.58 (Std. Dev.: 0.69). For Case #2, radiologist A rated it 4.17 (Std. Dev 0.75) from 2-D panels and 3.5 (Std. Dev 1.05) using the 3-D interactive tool. For the same case, radiologist B rated it 4.17 (Std. Dev. 1.13) from 2-D panels and rated it 3.67 (Std. Dev. 1.03) using the 3-D tool.

Conclusion: Six head/neck CT Angiography data sets were used to evaluate a bone segmentation algorithm. The segmented models were reviewed by 2 radiologists using 2-D static panels. They also used a VTK based tool that allowed minimal 3-D interactivity. There was no difference between the ratings obtained from the 2 radiologists for these cases, as well as using the 2 different evaluation techniques. (S.S., R.M. and V.K. are employees of GE Corporation; A.K. has a research agreement with GE.)




Questions about this event email: