An adaptive partitioning approach for mining discriminant regions in 3D image data

  • Authors:

  • Vasileios Megalooikonomou;Despina Kontos;Dragoljub Pokrajac;Aleksandar Lazarevic;Zoran Obradovic

  • Affiliations:

  • Data Engineering Laborarory (DEnLab), Temple University, Philadelphia, USA 19122-6094 and Center for Information Science and Technology, Temple University, Philadelphia, USA 19122-6094 and Departm ...;Data Engineering Laborarory (DEnLab), Temple University, Philadelphia, USA 19122-6094 and Center for Information Science and Technology, Temple University, Philadelphia, USA 19122-6094 and Departm ...;Computer and Information Science Department, Delaware State University, Dover, USA 19901 and Applied Mathematics Research Center (AMRC), Delaware State University, Dover, USA 19901;United Technologies Research Center, East Hartford, USA 06108;Center for Information Science and Technology, Temple University, Philadelphia, USA 19122-6094 and Department of Computer and Information Sciences, Temple University, Philadelphia, USA 19122-6094

  • Venue:
  • Journal of Intelligent Information Systems
  • Year:
  • 2008

Quantified Score

Hi-index 0.00

Visualization

Abstract

Mining discriminative spatial patterns in image data is an emerging subject of interest in medical imaging, meteorology, engineering, biology, and other fields. In this paper, we propose a novel approach for detecting spatial regions that are highly discriminative among different classes of three dimensional (3D) image data. The main idea of our approach is to treat the initial 3D image as a hyper-rectangle and search for discriminative regions by adaptively partitioning the space into progressively smaller hyper-rectangles (sub-regions). We use statistical information about each hyper-rectangle to guide the selectivity of the partitioning. A hyper-rectangle is partitioned only if its attribute cannot adequately discriminate among the distinct labeled classes, and it is sufficiently large for further splitting. To evaluate the discriminative power of the attributes corresponding to the detected regions, we performed classification experiments on artificial and real datasets. Our results show that the proposed method outperforms major competitors, achieving 30% and 15% better classification accuracy on synthetic and real data respectively while reducing by two orders of magnitude the number of statistical tests required by voxel-based approaches.