BRIEFCASE | SEARCH | EMAIL THIS EVENT | LOGIN    
RSNA 2003 Scientific Papers > Segmentation of Pulmonary Nodules with 3D Active Contour ...
 
  Scientific Papers
  SESSION: Physics (CAD VIII: Thoracic CT, Others)

Segmentation of Pulmonary Nodules with 3D Active Contour Model for Computer-aided Diagnosis

  DATE: Thursday, December 04 2003
  START TIME: 11:30 AM
  END TIME: 11:37 AM
  LOCATION: Room S401AB
  CODE: Q16-1342
TOOLS
 
ADD TO BRIEFCASE
  PRINT
  EMAIL

PARTICIPANTS
PRESENTER
Ted Way
Ann Arbor , MI
 
CO-AUTHOR
Berkman Sahiner PhD
 
Lubomir Hadjiiski PhD
 
Heang-Ping Chan PhD
 
Naama Bogot MD
 
Philip Cascade MD
 
et al
 

Keywords
Computed tomography (CT), image processing
Computers, diagnostic aid
Lung, CT
 
Abstract:
HTML

Purpose: To develop an automated lung nodule segmentation method for characterization of nodules on CT scans for computer-aided diagnosis (CAD).

Methods and Materials: We are developing 2D and 3D active contour (AC) models for nodule segmentation. The 3D AC model searches for the lesion boundary by minimizing a cost function with seven energy terms: (1) gradient, which rewards vertices on high gradients, (2) continuity, which encourages similar distances between vertices, (3) balloon, which expands or deflates the contour, (4) homogeneity, which seeks similar pixel intensities inside and outside the contour, (5) mask, which penalizes contours from growing into the lung wall, (6) curvature, which suppresses sharp angles between boundary segments, and (7) 3D curvature, which encourages continuity and smoothness between contours in the adjacent slices. The 2D AC model segments the nodule slice by slice using a cost function with the first six energy terms. In this preliminary study, the accuracy of the 2D and 3D AC models was compared by analyzing the segmented boundaries on a data set of 160 slices containing 26 nodules. Segmentation was scored based on visual assessment. A subjective score of 1 (excellent) through 5 (unacceptable) was given to each contour, based on the closeness of the segmented boundary to the perceived boundary and whether the boundary contained the lung wall for juxtapleural nodules or contained blood vessels for juxtavascular nodules.

Results: The 2D active contour resulted in 127 (79.8%) slices that were excellent, acceptable, or only needing slight improvement. The 3D model improved the nodule boundary on 48 (30%) slices, had no noticeable effect on 89 (55.6%) slices, but worsened the boundary on 23 (14.4%) slices. The number of slices rated as excellent was 57 (35.6%) and 43 (26.58%), for 3D and 2D segmentation, respectively. Degradation was mainly caused by faint nodules with low contrast and the overall cylindrical constraint imposed by the 3D curvature energy on the nodule.

Conclusion: Although the 2D AC model showed acceptable performance for most of the slices, our preliminary study indicates that the 2D model may be inadequate for tasks such as analysis of interval changes of nodule volume in serial exams that require excellent segmentation performance. The 3D AC model shows promise in improving the accuracy of lung nodule segmentation for classification in CAD. Further work is underway to optimize the energy terms and to evaluate the method with a larger dataset.

 

 

 


Questions about this event email: tway@eecs.umich.edu