Selecting Texture Measures for Detection of Pneumonia

N.M. Noor, O.M. Rijal, A. Yunus, S.A.R. Abu-Bakar, and G.C. Peng (Malaysia)


Pattern recognition, medicine, texture measures, andstatistical methods


This paper presents a statistical method for selecting texture measures that may be used for the detection of pneumonia patients and normal individuals when using digitized chest X-ray films. Two dimensional Daubechies wavelet features were obtained from the region of interest (ROI). Twelve texture measures were calculated yielding 48 descriptors or features for the purpose of recognizing pneumonia. Each ROI is then represented by a vector of texture measures which is transformed by the appropriate orthogonal matrices. The approximate confidence region for the first two elements of the transformed vectors is a first indicator of group separation for a given texture measure. These bivariate vectors were found to be normally distributed and hence the Linear Discrimination Function and Quadratic Discrimination Function were estimated. Using both the approximate confidence region and the discriminant function, estimates of the probability of Type I and Type II errors were obtained using a second data set. The results showed the probability of both errors were smaller when the discriminant function was used relative to the approximate confidence region. Finally mean value, maximum row sum energy, maximum column sum energy and combination of twelve features using second orthogonal transformation matrix gives the lowest Type II error.

Important Links:

Go Back