Surface Curvature Line Clustering for Polyp Detection in CT Colonography

Abstract

Automatic polyp detection is a helpful addition to laborious visual inspection in CT colonography. Traditional detection methods are based on calculating image features at discrete positions on the colon wall. However large-scale surface shapes are not captured. This paper presents a novel approach to aggregate surface shape information for automatic polyp detection. The iso-surface of the colon wall can be partitioned into geometrically homogeneous regions based on clustering of curvature lines, using a spectral clustering algorithm and a symmetric line similarity measure. Each partition corresponds with the surface area that is covered by a single cluster. For each of the clusters, a number of features are calculated, based on the volumetric shape index and the surface curvedness, to select the surface partition corresponding to the cap of a polyp. We have applied our clustering approach to nine annotated patient datasets. Results show that the surface partition-based features are highly correlated with true polyp detections and can thus be used to reduce the number of false-positive detections.