ROC evaluation of statistical wavelet-based analysis of brain activation in [15O]-H2O PET scans

Neuroimage. 2005 Feb 1;24(3):763-70. doi: 10.1016/j.neuroimage.2004.08.052.

Abstract

This paper presents and evaluates a wavelet-based statistical analysis of PET images for the detection of brain activation areas. Brain regions showing significant activations were obtained by performing Student's t tests in the wavelet domain, reconstructing the final image from only those wavelet coefficients that passed the statistical test at a given significance level, and discarding artifacts introduced during the reconstruction process. Using Receiver Operating Characteristic (ROC) curves, we have compared this statistical analysis in the wavelet domain to the conventional image-domain Statistical Parametric Mapping (SPM) method. For obtaining an accurate assessment of sensitivity and specificity, we have simulated realistic single subject [15O]-H2O PET studies with different hyperactivation levels of the thalamic region. The results obtained from an ROC analysis show that the wavelet approach outperforms conventional SPM in identifying brain activation patterns. Using the wavelet method, activation areas detected were closer in size and shape to the region actually activated in the reference image.

Publication types

  • Clinical Trial
  • Research Support, Non-U.S. Gov't

MeSH terms

  • Algorithms
  • Brain / diagnostic imaging*
  • Brain / physiology*
  • Brain Mapping
  • False Positive Reactions
  • Humans
  • Image Interpretation, Computer-Assisted
  • Magnetic Resonance Imaging
  • Oxygen Radioisotopes
  • Positron-Emission Tomography / statistics & numerical data*
  • ROC Curve

Substances

  • Oxygen Radioisotopes