A New GLLD Operator for Mass Detection in Digital Mammograms

Abstract

During the last decade, several works have dealt with computer automatic diagnosis (CAD) of masses in digital mammograms. Generally, the main difficulty remains the detection of masses. This work proposes an efficient methodology for mass detection based on a new local feature extraction. Local binary pattern (LBP) operator and its variants proposed by Ojala are a powerful tool for textures classification. However, it has been proved that such operators are not able to model at their own texture masses. We propose in this paper a new local pattern model named gray level and local difference (GLLD) where we take into consideration absolute gray level values as well as local difference as local binary features. Artificial neural networks (ANNs), support vector machine (SVM), and k-nearest neighbors (kNNs) are, then, used for classifying masses from nonmasses, illustrating better performance of ANN classifier. We have used 1000 regions of interest (ROIs) obtained from the Digital Database for Screening Mammography (DDSM). The area under the curve of the corresponding approach has been found to be A(z) = 0.95 for the mass detection step. A comparative study with previous approaches proves that our approach offers the best performances.

Figure 1

Example of basic LBP operator.

Figure 2

The central pixel g _c and its P circularly symmetric neighbor with radius R.

Figure 3

(a) A 3∗3 block with central pixel corresponding to the mean value of its neighbors. (b) The sign components. (c) The magnitude components.

Figure 4

Different processing steps of of the proposed GLLD based approach.

Figure 5

The image results after the application of the three operator and their fusion.

Figure 6

The extraction of the different local features from an ROI sample. Step (I): the texture features can be computed by building the histogram over the corresponding ROI. Step (II): the histogram from the three operators is concatenated to build the texture features of the selected ROI.

Figure 7

GLLD feature distributions extracted and concatenated to constitute the final histogram.

Figure 8

MLP classifier architecture.

Figure 9

Implementation of the proposed method.

Figure 10

ROC curve corresponding to a subset of 1000 ROIs images from the DDSM database.