Towards feature selection for digital mammogram classification

Extended Mammogram Classification From Textural Features

“…In our research, we extended the approach presented in [2] in order to be able to apply it to other mammogram datasets. The system in [2] is not working accordingly if the breast tissue has too low intensity or if white patches appear on the image.…”

“…A multi-class classifier is built to distinguish normal, benign and malignant classes. We start from a recently proposed CAD system [2], which consists of the following five steps:…”

“…In reference [2], multiple experiments are reported with competitive results for medical images from the MIAS database [15] using Principal Component Analysis (PCA) for feature selection and Random Forest (RF) classifiers. As the system was created and tested for MIAS images, our main objective in this study is to determine if the system can be extended to obtain a good performance for other medical images.…”

“…Considering the abovementioned issues, in the current paper we extend the previously proposed system mainly concerning the preprocessing and feature extraction steps. The preprocessing step is extended by (1) removing white patches outside the breast, (2) setting the threshold to define the breast dynamically, instead of hard-coding it, (3) creating a more robust method to define the location of the pectoral muscle, (4) selecting the proper component overlapping with the pectoral muscle and (5) filling its concave parts.…”

“…Moreover, feature extraction considers characteristics calculated from Gray-Level Co-Occurrence Matrices (GLCM) with different distances and is reimplemented to use the graphical processing unit. For segmentation, feature selection, and classification, the same methods are used as in [2], namely, kmeans, PCA and genetic algorithm (GA) based feature selection, and Decision Tree (DT) and RF respectively.…”