Structural feature selection for connectivity network-based MCI diagnosis

  • Authors:

  • Biao Jie;Daoqiang Zhang;Chong-Yaw Wee;Dinggang Shen

  • Affiliations:

  • Dept. of Computer Science and Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing, China, Dept. of Radiology and BRIC, University of North Carolina at Chapel Hill, NC;Dept. of Computer Science and Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing, China, Dept. of Radiology and BRIC, University of North Carolina at Chapel Hill, NC;Dept. of Radiology and BRIC, University of North Carolina at Chapel Hill, NC;Dept. of Radiology and BRIC, University of North Carolina at Chapel Hill, NC

  • Venue:
  • MBIA'12 Proceedings of the Second international conference on Multimodal Brain Image Analysis
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

Connectivity networks have been recently used for classification of neurodegenerative diseases, e.g., mild cognitive impairment (MCI). In typical connectivity network-based classification, features are often extracted from (multiple) connectivity networks and concatenated into a long vector for subsequent feature selection and classification. However, some useful network topological information may be lost in this type of approach. In this paper, we propose a new structural feature selection method which embeds the topological information of connectivity networks through graph kernel and then uses recursive feature elimination with graph kernel (RFE-GK) to select the most discriminative features. Furthermore, multiple kernel learning (MKL) is also adopted to combine multiple graph kernels for joint structural feature selectionfrom multiple connectivity networks. The experimental results show the efficacy of our proposed method with comparison to the state-of-the-art method in MCI classification, based on the connectivity networks.