A Sparse-modeled ROI for GLAM Construction in Image Classification Problems—A Case Study of Breast Cancer - Computational Intelligence Techniques for Comparative Genomics - page 45

adopted by the peer scientists [ 6 , 7 ]. This novel idea has drastically reduced time
complexity in constructing GLAM. The organization of the rest of the paper is as
follows. Section 2 includes the motivation. Section 3 illustrates the methodology of
sparse-modeled ROI. Sections 4 and 5 report experimental observations and
conclusions.

2 Motivation

Many computer-aided classi cation systems are proposed for classi cation of

masses of mammograms [ 8 ]. Identi cation of ROI using segmentation is the pre-

liminary phase in the process of classi

cation [ 9 , 10 ]. Feature extraction and

classi cation [ 11 ]. Textural

feature extraction from ROI using gray level statistical matrices is the key issue in
most of these systems [ 12 ]. The Gray Level Aura Matrix (GLAM) is a statistical
method for extracting texture information from the images. The GLAM charac-
terizes the spatial distribution of gray levels [ 13 ]. The size and shape of ROIs
depend on the size and shape of suspected regions segmented in the preliminary
phase. The size of ROI not only in

cation are the lateral steps of mammogram classi

uences the computational time to construct the
statistical matrix but also the classication rate. The scientists are choosing rect-
angular-shaped window for ROI. They are selecting different window sizes
depending on their study and database that

fl

they have selected. For example,

Chandy et al. [ 14 ] selected 200 × 200 pixels as uniform size of ROI. Mohanty et al.

[ 15 ] 50 pixels for ROI size for DDSM databases. In another study,

Mohanty et al. [ 16 ] considered 256

xed 50 ×

× 256 pixels window for classi cation and

detection of breast cancer. Hussain [ 17 ] used ROIs of size from 267 × 274 pixels to

1,197 × 1,301 pixels depending on size of the mass. For classi cation of mass, the

authors have selected 1,024 1,024 pixels form mass center [ 18 ]. The regular

rectangular shape and uniform size for all ROIs include more number of unwanted
pixels which increases the computational time to construct gray level matrices, a
feature extraction process. For example, in the Figs. 1 , 2 and 3 , no ROI is of
rectangular shape. Total number of pixels of ROI of the

×

rst image is 2,191, second
is 839 and third contains 68 pixels only. Therefore, we have proposed a novel idea
not to consider the pixels outside the arbitrary shape of mass region. The model and
algorithm explained in the next section.

(a) (b) (c) (d)

Fig. 1 Original mammograms